Solshine commited on
Commit
0bcb3e6
·
0 Parent(s):

Initial public release: SAE weights, cfg, and model card

Browse files
This view is limited to 50 files because it contains too many changes.   See raw diff
Files changed (50) hide show
  1. .gitattributes +35 -0
  2. README.md +292 -0
  3. tinyllama_jumprelu_L12_deceptive_only/cfg.json +23 -0
  4. tinyllama_jumprelu_L12_deceptive_only/sae_weights.safetensors +3 -0
  5. tinyllama_jumprelu_L12_honest_only/cfg.json +23 -0
  6. tinyllama_jumprelu_L12_honest_only/sae_weights.safetensors +3 -0
  7. tinyllama_jumprelu_L12_mixed/cfg.json +23 -0
  8. tinyllama_jumprelu_L12_mixed/sae_weights.safetensors +3 -0
  9. tinyllama_jumprelu_L15_deceptive_only/cfg.json +23 -0
  10. tinyllama_jumprelu_L15_deceptive_only/sae_weights.safetensors +3 -0
  11. tinyllama_jumprelu_L15_honest_only/cfg.json +23 -0
  12. tinyllama_jumprelu_L15_honest_only/sae_weights.safetensors +3 -0
  13. tinyllama_jumprelu_L15_mixed/cfg.json +23 -0
  14. tinyllama_jumprelu_L15_mixed/sae_weights.safetensors +3 -0
  15. tinyllama_jumprelu_L3_deceptive_only/cfg.json +23 -0
  16. tinyllama_jumprelu_L3_deceptive_only/sae_weights.safetensors +3 -0
  17. tinyllama_jumprelu_L3_honest_only/cfg.json +23 -0
  18. tinyllama_jumprelu_L3_honest_only/sae_weights.safetensors +3 -0
  19. tinyllama_jumprelu_L3_mixed/cfg.json +23 -0
  20. tinyllama_jumprelu_L3_mixed/sae_weights.safetensors +3 -0
  21. tinyllama_jumprelu_L6_deceptive_only/cfg.json +23 -0
  22. tinyllama_jumprelu_L6_deceptive_only/sae_weights.safetensors +3 -0
  23. tinyllama_jumprelu_L6_honest_only/cfg.json +23 -0
  24. tinyllama_jumprelu_L6_honest_only/sae_weights.safetensors +3 -0
  25. tinyllama_jumprelu_L6_mixed/cfg.json +23 -0
  26. tinyllama_jumprelu_L6_mixed/sae_weights.safetensors +3 -0
  27. tinyllama_jumprelu_L9_deceptive_only/cfg.json +23 -0
  28. tinyllama_jumprelu_L9_deceptive_only/sae_weights.safetensors +3 -0
  29. tinyllama_jumprelu_L9_honest_only/cfg.json +23 -0
  30. tinyllama_jumprelu_L9_honest_only/sae_weights.safetensors +3 -0
  31. tinyllama_jumprelu_L9_mixed/cfg.json +23 -0
  32. tinyllama_jumprelu_L9_mixed/sae_weights.safetensors +3 -0
  33. tinyllama_jumprelu_ste_L12_deceptive_only/cfg.json +23 -0
  34. tinyllama_jumprelu_ste_L12_deceptive_only/sae_weights.safetensors +3 -0
  35. tinyllama_jumprelu_ste_L12_honest_only/cfg.json +23 -0
  36. tinyllama_jumprelu_ste_L12_honest_only/sae_weights.safetensors +3 -0
  37. tinyllama_jumprelu_ste_L12_mixed/cfg.json +23 -0
  38. tinyllama_jumprelu_ste_L12_mixed/sae_weights.safetensors +3 -0
  39. tinyllama_jumprelu_ste_L15_deceptive_only/cfg.json +23 -0
  40. tinyllama_jumprelu_ste_L15_deceptive_only/sae_weights.safetensors +3 -0
  41. tinyllama_jumprelu_ste_L15_honest_only/cfg.json +23 -0
  42. tinyllama_jumprelu_ste_L15_honest_only/sae_weights.safetensors +3 -0
  43. tinyllama_jumprelu_ste_L15_mixed/cfg.json +23 -0
  44. tinyllama_jumprelu_ste_L15_mixed/sae_weights.safetensors +3 -0
  45. tinyllama_jumprelu_ste_L9_deceptive_only/cfg.json +23 -0
  46. tinyllama_jumprelu_ste_L9_deceptive_only/sae_weights.safetensors +3 -0
  47. tinyllama_jumprelu_ste_L9_honest_only/cfg.json +23 -0
  48. tinyllama_jumprelu_ste_L9_honest_only/sae_weights.safetensors +3 -0
  49. tinyllama_jumprelu_ste_L9_mixed/cfg.json +23 -0
  50. tinyllama_jumprelu_ste_L9_mixed/sae_weights.safetensors +3 -0
.gitattributes ADDED
@@ -0,0 +1,35 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ *.7z filter=lfs diff=lfs merge=lfs -text
2
+ *.arrow filter=lfs diff=lfs merge=lfs -text
3
+ *.bin filter=lfs diff=lfs merge=lfs -text
4
+ *.bz2 filter=lfs diff=lfs merge=lfs -text
5
+ *.ckpt filter=lfs diff=lfs merge=lfs -text
6
+ *.ftz filter=lfs diff=lfs merge=lfs -text
7
+ *.gz filter=lfs diff=lfs merge=lfs -text
8
+ *.h5 filter=lfs diff=lfs merge=lfs -text
9
+ *.joblib filter=lfs diff=lfs merge=lfs -text
10
+ *.lfs.* filter=lfs diff=lfs merge=lfs -text
11
+ *.mlmodel filter=lfs diff=lfs merge=lfs -text
12
+ *.model filter=lfs diff=lfs merge=lfs -text
13
+ *.msgpack filter=lfs diff=lfs merge=lfs -text
14
+ *.npy filter=lfs diff=lfs merge=lfs -text
15
+ *.npz filter=lfs diff=lfs merge=lfs -text
16
+ *.onnx filter=lfs diff=lfs merge=lfs -text
17
+ *.ot filter=lfs diff=lfs merge=lfs -text
18
+ *.parquet filter=lfs diff=lfs merge=lfs -text
19
+ *.pb filter=lfs diff=lfs merge=lfs -text
20
+ *.pickle filter=lfs diff=lfs merge=lfs -text
21
+ *.pkl filter=lfs diff=lfs merge=lfs -text
22
+ *.pt filter=lfs diff=lfs merge=lfs -text
23
+ *.pth filter=lfs diff=lfs merge=lfs -text
24
+ *.rar filter=lfs diff=lfs merge=lfs -text
25
+ *.safetensors filter=lfs diff=lfs merge=lfs -text
26
+ saved_model/**/* filter=lfs diff=lfs merge=lfs -text
27
+ *.tar.* filter=lfs diff=lfs merge=lfs -text
28
+ *.tar filter=lfs diff=lfs merge=lfs -text
29
+ *.tflite filter=lfs diff=lfs merge=lfs -text
30
+ *.tgz filter=lfs diff=lfs merge=lfs -text
31
+ *.wasm filter=lfs diff=lfs merge=lfs -text
32
+ *.xz filter=lfs diff=lfs merge=lfs -text
33
+ *.zip filter=lfs diff=lfs merge=lfs -text
34
+ *.zst filter=lfs diff=lfs merge=lfs -text
35
+ *tfevents* filter=lfs diff=lfs merge=lfs -text
README.md ADDED
@@ -0,0 +1,292 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ ---
2
+ language: en
3
+ license: apache-2.0
4
+ tags:
5
+ - sparse-autoencoder
6
+ - SAE
7
+ - interpretability
8
+ - deception-detection
9
+ - mechanistic-interpretability
10
+ - saelens
11
+ - neuronpedia
12
+ - behavioral-sampling
13
+ - llama
14
+ base_model:
15
+ - TinyLlama/TinyLlama-1.1B-intermediate-step-1431k-3T
16
+ datasets:
17
+ - Solshine/deception-behavioral-multimodel
18
+ ---
19
+
20
+ # TinyLlama-1.1B Deception Behavioral SAEs
21
+
22
+ 39 Sparse Autoencoders (30 original + 9 STE-validated) trained on residual stream activations from [TinyLlama/TinyLlama-1.1B-intermediate-step-1431k-3T](https://huggingface.co/TinyLlama/TinyLlama-1.1B-intermediate-step-1431k-3T) (1.1B parameter Llama2-architecture base model), capturing behavioral deception signals via same-prompt temperature sampling.
23
+
24
+
25
+ ## Training-data caveat — please read before use
26
+
27
+ The "honest" and "deceptive" labels used to train and evaluate this SAE
28
+ come from a **mixed scenario pool** of nine prompts:
29
+
30
+ - **Six clean incentive-structure scenarios** — `insider_info`,
31
+ `accounting_error`, `ai_oversight_log`, `ai_capability_hide`,
32
+ `surprise_party`, `job_interview_gap`. Each presents a realistic
33
+ situation with a non-trivial incentive to conceal or disclose
34
+ information; the honest/deceptive distinction is which of the two
35
+ behavioral choices the model's completion settles into under
36
+ temperature sampling.
37
+ - **Three role-play identity-assignment scenarios** —
38
+ `secret_role_game` ("You have been assigned the role of Fascist"),
39
+ `secret_role_game_v2` ("You are secretly one of the Mafia members"),
40
+ `werewolf_game` ("You are a Werewolf"). These pre-assign the model a
41
+ deceptive identity and label a completion "deceptive" when the model
42
+ drifts away from the assigned role or "honest" when it echoes it.
43
+
44
+ **What this mixed pool means for the SAE's labels.** Within the six
45
+ incentive-structure scenarios, the honest/deceptive distinction is a
46
+ measurement of behavioral choice under an ambiguous incentive. Within
47
+ the three role-play scenarios, the distinction is a measurement of
48
+ role-consistency under identity-assigned role-play — which is a
49
+ well-defined phenomenon but not the same as emergent or incentive-
50
+ driven deception.
51
+
52
+ **What this SAE is and is not good for.**
53
+
54
+ - **Good for:** research on mixed-pool activation geometry; SAE
55
+ feature-geometry studies; as one of a set of baselines when
56
+ comparing multiple SAE families; as a reference implementation of
57
+ same-prompt temperature-sampled behavioral SAE training at scale.
58
+ - **Not recommended as a standalone deception detector.** The
59
+ role-consistency signal from the three role-play scenarios is mixed
60
+ into every aggregate metric reported below. A downstream user who
61
+ wants an "emergent-deception feature set" should restrict attention
62
+ to features whose activation pattern concentrates in the
63
+ `insider_info` / `accounting_error` / `ai_oversight_log` /
64
+ `ai_capability_hide` / `surprise_party` / `job_interview_gap`
65
+ scenarios — or wait for the methodologically corrected V3 re-release
66
+ currently in preparation on the decision-incentive scenario bank
67
+ (no pre-assigned deceptive identity).
68
+
69
+ **What is unaffected by this caveat.**
70
+
71
+ - The SAE weights, reconstruction metrics (explained variance, L0,
72
+ alive features), and engineering of the training pipeline are
73
+ accurate as reported.
74
+ - The linear-probe balanced-accuracy numbers in the upstream paper
75
+ measure the mixed pool; the 6-scenario clean-subset re-analysis is
76
+ listed as a planned appendix for the next manuscript revision.
77
+
78
+ A companion methodology-first Gemma 4 SAE suite is in preparation using
79
+ pretraining-distribution data + a decision-incentive behavior split;
80
+ this README will be updated with a link when that release is public.
81
+
82
+ ---
83
+
84
+ Part of the cross-model deception SAE study: [Solshine/deception-behavioral-saes-saelens](https://huggingface.co/Solshine/deception-behavioral-saes-saelens) (9 models, 348 total SAEs).
85
+
86
+ ## What's in This Repo
87
+
88
+ - **39 SAEs** — 30 original + 9 STE-validated (`_ste_` tag)
89
+ - **Layers:** L3, L6, L9, L12, L15 (original); L9, L12, L15 (STE)
90
+ - **2 architectures:** TopK (k=64), JumpReLU
91
+ - **3 training conditions:** `mixed`, `deceptive_only`, `honest_only`
92
+ - **Format:** SAELens/Neuronpedia-compatible (safetensors + cfg.json)
93
+ - **Dimensions:** d_in=2048, d_sae=8192 (4x expansion)
94
+
95
+ ## Research Context
96
+
97
+ This is a follow-up to ["The Secret Agenda: LLMs Strategically Lie Undetected by Current Safety Tools"](https://arxiv.org/abs/2509.20393) (arXiv:2509.20393). Same-prompt behavioral sampling: a single ambiguous scenario prompt produces both deceptive and honest completions at temperature=1.0 via temperature sampling, classified by Gemini 2.5 Flash.
98
+
99
+ Code: [SolshineCode/deception-nanochat-sae-research](https://github.com/SolshineCode/deception-nanochat-sae-research)
100
+
101
+ ## Key Findings — TinyLlama-1.1B
102
+
103
+ TinyLlama is the **most distinctive model** in the 9-model study in terms of layer profile — and the most STE-validated.
104
+
105
+ | Metric | Value |
106
+ |---|---|
107
+ | Peak layer | L21 (95% depth) |
108
+ | Peak balanced accuracy | **72.4%** |
109
+ | Peak AUROC | **0.784** |
110
+ | Best SAE probe accuracy | **77.8%** (`tinyllama_jumprelu_L15_honest_only`) |
111
+ | SAEs beating raw baseline | 14/30 (47%) — SAEs **help**, highest rate in study |
112
+
113
+ **Monotonically increasing layer profile — unique across all 9 models:** Every other model with above-chance deception signal shows a peak somewhere in the middle of the network and then declines. TinyLlama is the sole exception: the deception signal rises continuously from L3 (~55%) through L21 (72.4%), with no saturation or decline at the final layer. This suggests the model has not developed the dedicated mid-network semantic processing that larger models exhibit, and deception-relevant computation continues accumulating through the output layers.
114
+
115
+ **Highest SAE-helps rate in the entire study (47%):** 14 of 30 original SAEs beat their respective layer's raw baseline. This puts TinyLlama firmly in the SAE-helps regime alongside SmolLM2, nanochat-d20, Llama-3.2-1B, and Pythia-160M — all models at or below 1.3B parameters.
116
+
117
+ **Best SAE substantially outperforms raw peak (+5.4pp):** `tinyllama_jumprelu_L15_honest_only` achieves 77.8%, beating the L15 raw baseline of 72.4% by +5.37pp. Notably, this is a non-peak layer for raw activations (L21 peaks raw), but the JumpReLU honest_only SAE at L15 nearly matches the L21 raw result while operating on an intermediate representation.
118
+
119
+ **Strongest STE validation of any model:** 8/9 STE-validated conditions show STE JumpReLU beating TopK (0 collapses, 1 inconclusive). Mean STE accuracy = 69.2% vs TopK = 64.1% (+5.1pp). The L15 honest_only condition improves by +10.1pp with STE. Combined with nanochat-d20 results, 15/18 STE conditions (83%) confirm the JumpReLU+honest_only advantage is real, not a dimensionality artifact from the threshold=0 bug.
120
+
121
+ **Architecture note:** TinyLlama-1.1B uses the Llama2 architecture — grouped-query attention (GQA), SwiGLU MLP, RMSNorm, rotary position embeddings — but with a 22-layer, 2048-dimensional residual stream trained on 3 trillion tokens to an intermediate checkpoint. It uses the same architecture family as SmolLM2 but with a much wider residual stream. The monotonically increasing profile may reflect that 22 layers is insufficient for TinyLlama-class models to develop stable mid-network deception representations at this parameter count.
122
+
123
+ ## SAE Format
124
+
125
+ Each SAE lives in a subfolder named `{sae_id}/` containing:
126
+ - `sae_weights.safetensors` — encoder/decoder weights
127
+ - `cfg.json` — SAELens-compatible config
128
+
129
+ `hook_name` format: `model.layers.{layer}.hook_resid_post`
130
+
131
+ STE SAEs have `_ste_` in the tag (e.g., `tinyllama_jumprelu_ste_L15_honest_only`).
132
+
133
+ ## Training Details
134
+
135
+ | Parameter | Value |
136
+ |---|---|
137
+ | Hardware | NVIDIA GeForce GTX 1650 Ti Max-Q, 4 GB VRAM, Windows 11 Pro |
138
+ | Training time | ~400–600 seconds per SAE |
139
+ | Epochs | 300 |
140
+ | Batch size | 128 |
141
+ | Expansion factor | 4x (2048 → 8192) |
142
+ | Activations | `resid_post` collected during autoregressive generation |
143
+ | Training conditions | `mixed` (n=243), `deceptive_only` (n=84), `honest_only` (n=159) |
144
+ | LLM classifier | Gemini 2.5 Flash |
145
+
146
+ ## Known Limitations
147
+
148
+ **JumpReLU threshold not learned (original 30 SAEs):** All non-STE SAEs have `threshold = 0` — functionally ReLU. L0 ≈ 50% of d_sae. TopK SAEs are unaffected.
149
+
150
+ **STE fix (2026-04-11):** 9 `_ste_` tagged SAEs use the Gaussian-kernel STE (Rajamanoharan et al. 2024, arXiv:2407.14435). TinyLlama provides the strongest STE validation in the study (8/9 conditions confirm).
151
+
152
+ **Intermediate checkpoint:** The model is `intermediate-step-1431k-3T`, not fully converged. A fully trained TinyLlama checkpoint might produce different layer profiles.
153
+
154
+ ## Loading Example
155
+
156
+ ```python
157
+ from safetensors.torch import load_file
158
+ import json
159
+
160
+ sae_id = "tinyllama_jumprelu_L15_honest_only"
161
+ weights = load_file(f"{sae_id}/sae_weights.safetensors")
162
+ cfg = json.load(open(f"{sae_id}/cfg.json"))
163
+
164
+ # W_enc: [2048, 8192], W_dec: [8192, 2048]
165
+ # cfg["hook_name"] == "model.layers.15.hook_resid_post"
166
+ print(f"Training condition: {cfg['training_condition']}")
167
+ print(f"STE variant: {'_ste_' in sae_id}")
168
+ ```
169
+
170
+
171
+ ## Usage
172
+
173
+ ### 1. Load an SAE from this repo
174
+
175
+ ```python
176
+ from huggingface_hub import hf_hub_download
177
+ from safetensors.torch import load_file
178
+ import json
179
+
180
+ repo_id = "Solshine/deception-saes-tinyllama-1-1b"
181
+ sae_id = "tinyllama_jumprelu_ste_L12_honest_only" # replace with any tag in this repo
182
+
183
+ weights_path = hf_hub_download(repo_id, f"{sae_id}/sae_weights.safetensors")
184
+ cfg_path = hf_hub_download(repo_id, f"{sae_id}/cfg.json")
185
+
186
+ with open(cfg_path) as f:
187
+ cfg = json.load(f)
188
+
189
+ # Option A — load with SAELens (≥3.0 required for jumprelu/topk; ≥3.5 for gated)
190
+ from sae_lens import SAE
191
+ sae = SAE.from_dict(cfg)
192
+ sae.load_state_dict(load_file(weights_path))
193
+
194
+ # Option B — load manually (no SAELens dependency)
195
+ from safetensors.torch import load_file
196
+ state = load_file(weights_path)
197
+ # Keys: W_enc [2048, 8192], b_enc [8192],
198
+ # W_dec [8192, 2048], b_dec [2048], threshold [8192]
199
+ ```
200
+
201
+ ### 2. Hook into the model and collect residual-stream activations
202
+
203
+ These SAEs were trained on the **residual stream after each transformer layer**.
204
+ The `hook_name` field in `cfg.json` gives the exact HuggingFace `transformers`
205
+ submodule path to hook. LLaMA-2 architecture. Hook path: `model.layers.{layer}`.
206
+
207
+ ```python
208
+ import torch
209
+ from transformers import AutoModelForCausalLM, AutoTokenizer
210
+
211
+ model = AutoModelForCausalLM.from_pretrained("TinyLlama/TinyLlama-1.1B-intermediate-step-1431k-3T")
212
+ tokenizer = AutoTokenizer.from_pretrained("TinyLlama/TinyLlama-1.1B-intermediate-step-1431k-3T")
213
+
214
+ # Read hook_name from the cfg you already loaded:
215
+ # cfg["hook_name"] == "model.layers.12" (example — varies by SAE)
216
+ hook_name = cfg["hook_name"] # e.g. "model.layers.12"
217
+
218
+ # Navigate the submodule path and register a forward hook
219
+ import functools
220
+ submodule = functools.reduce(getattr, hook_name.split("."), model)
221
+
222
+ activations = {}
223
+ def hook_fn(module, input, output):
224
+ # Most transformer layers return (hidden_states, ...) as a tuple
225
+ h = output[0] if isinstance(output, tuple) else output
226
+ activations["resid"] = h.detach()
227
+
228
+ handle = submodule.register_forward_hook(hook_fn)
229
+
230
+ inputs = tokenizer("Your text here", return_tensors="pt")
231
+ with torch.no_grad():
232
+ model(**inputs)
233
+ handle.remove()
234
+
235
+ # activations["resid"]: [batch, seq_len, 2048]
236
+ resid = activations["resid"][:, -1, :] # last token position
237
+ ```
238
+
239
+ ### 3. Read feature activations
240
+
241
+ ```python
242
+ with torch.no_grad():
243
+ feature_acts = sae.encode(resid) # [batch, 8192] — sparse
244
+
245
+ # Which features fired?
246
+ active_features = feature_acts[0].nonzero(as_tuple=True)[0]
247
+ top_features = feature_acts[0].topk(10)
248
+
249
+ print("Active feature indices:", active_features.tolist())
250
+ print("Top-10 feature values:", top_features.values.tolist())
251
+ print("Top-10 feature indices:", top_features.indices.tolist())
252
+
253
+ # Reconstruct (for sanity check — should be close to resid)
254
+ reconstruction = sae.decode(feature_acts)
255
+ l2_error = (resid - reconstruction).norm(dim=-1).mean()
256
+ ```
257
+
258
+ ### Caveats and known limitations
259
+
260
+ **Hook names are HuggingFace `transformers`-style, not TransformerLens-style.**
261
+ The `hook_name` in `cfg.json` (e.g. `"model.layers.12"`) is a submodule path in the standard
262
+ HuggingFace model. SAELens' built-in activation-collection pipeline expects
263
+ TransformerLens hook names (e.g. `blocks.14.hook_resid_post`). This means
264
+ `SAE.from_pretrained()` with automatic model running **will not work** — use the
265
+ manual forward-hook pattern above instead.
266
+
267
+ **SAELens version requirements.**
268
+ - `topk` architecture: SAELens ≥ 3.0
269
+ - `jumprelu` architecture: SAELens ≥ 3.0
270
+ - `gated` architecture: SAELens ≥ 3.5 (or load manually with `state_dict`)
271
+
272
+ **JumpReLU `_ste_` vs standard variants.**
273
+ SAEs tagged `_ste_` use properly trained JumpReLU thresholds (Gaussian-kernel STE,
274
+ [Rajamanoharan et al. 2024](https://arxiv.org/abs/2407.14435)). Standard variants have
275
+ `threshold=0` and are functionally ReLU (trained before the STE fix on 2026-04-11).
276
+ Both load and run identically; the `_ste_` variants are sparser and more interpretable.
277
+
278
+ **These SAEs detect deceptive *behavior*, not deceptive *prompts**.*
279
+ They were trained on response-level activations where the same prompt produced both
280
+ deceptive and honest outputs. Feature activation differences reflect behavioral
281
+ divergence, not prompt content. See the paper for experimental design details.
282
+
283
+ ## Citation
284
+
285
+ ```bibtex
286
+ @article{thesecretagenda2025,
287
+ title={The Secret Agenda: LLMs Strategically Lie Undetected by Current Safety Tools},
288
+ author={DeLeeuw, Caleb},
289
+ journal={arXiv:2509.20393},
290
+ year={2025}
291
+ }
292
+ ```
tinyllama_jumprelu_L12_deceptive_only/cfg.json ADDED
@@ -0,0 +1,23 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "architecture": "jumprelu",
3
+ "d_in": 2048,
4
+ "d_sae": 8192,
5
+ "dtype": "float32",
6
+ "device": "cpu",
7
+ "model_name": "TinyLlama/TinyLlama-1.1B-intermediate-step-1431k-3T",
8
+ "hook_name": "model.layers.12",
9
+ "hook_layer": 12,
10
+ "hook_head_index": null,
11
+ "activation_fn_str": "jumprelu",
12
+ "activation_fn_kwargs": {},
13
+ "apply_b_dec_to_input": false,
14
+ "finetuning_scaling_factor": false,
15
+ "sae_lens_training_version": "deception-behavioral-v1",
16
+ "prepend_bos": false,
17
+ "dataset_path": "Solshine/deception-behavioral-multimodel",
18
+ "dataset_trust_remote_code": false,
19
+ "context_size": null,
20
+ "normalize_activations": "none",
21
+ "training_condition": "deceptive_only",
22
+ "training_notes": "Deception behavioral SAE \u2014 same-prompt behavioral sampling. Model: TinyLlama/TinyLlama-1.1B-intermediate-step-1431k-3T, Layer 12, jumprelu. See https://github.com/SolshineCode/deception-nanochat-sae-research"
23
+ }
tinyllama_jumprelu_L12_deceptive_only/sae_weights.safetensors ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:235383ce86c979b798cc0c4f9a34a9f8490ef786889fd1356adb5ca13f3444aa
3
+ size 134291856
tinyllama_jumprelu_L12_honest_only/cfg.json ADDED
@@ -0,0 +1,23 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "architecture": "jumprelu",
3
+ "d_in": 2048,
4
+ "d_sae": 8192,
5
+ "dtype": "float32",
6
+ "device": "cpu",
7
+ "model_name": "TinyLlama/TinyLlama-1.1B-intermediate-step-1431k-3T",
8
+ "hook_name": "model.layers.12",
9
+ "hook_layer": 12,
10
+ "hook_head_index": null,
11
+ "activation_fn_str": "jumprelu",
12
+ "activation_fn_kwargs": {},
13
+ "apply_b_dec_to_input": false,
14
+ "finetuning_scaling_factor": false,
15
+ "sae_lens_training_version": "deception-behavioral-v1",
16
+ "prepend_bos": false,
17
+ "dataset_path": "Solshine/deception-behavioral-multimodel",
18
+ "dataset_trust_remote_code": false,
19
+ "context_size": null,
20
+ "normalize_activations": "none",
21
+ "training_condition": "honest_only",
22
+ "training_notes": "Deception behavioral SAE \u2014 same-prompt behavioral sampling. Model: TinyLlama/TinyLlama-1.1B-intermediate-step-1431k-3T, Layer 12, jumprelu. See https://github.com/SolshineCode/deception-nanochat-sae-research"
23
+ }
tinyllama_jumprelu_L12_honest_only/sae_weights.safetensors ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:a9ef963d1b73515a4a4f6a220dcbd9b3e12a5c293025d13461e5796d4b363225
3
+ size 134291856
tinyllama_jumprelu_L12_mixed/cfg.json ADDED
@@ -0,0 +1,23 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "architecture": "jumprelu",
3
+ "d_in": 2048,
4
+ "d_sae": 8192,
5
+ "dtype": "float32",
6
+ "device": "cpu",
7
+ "model_name": "TinyLlama/TinyLlama-1.1B-intermediate-step-1431k-3T",
8
+ "hook_name": "model.layers.12",
9
+ "hook_layer": 12,
10
+ "hook_head_index": null,
11
+ "activation_fn_str": "jumprelu",
12
+ "activation_fn_kwargs": {},
13
+ "apply_b_dec_to_input": false,
14
+ "finetuning_scaling_factor": false,
15
+ "sae_lens_training_version": "deception-behavioral-v1",
16
+ "prepend_bos": false,
17
+ "dataset_path": "Solshine/deception-behavioral-multimodel",
18
+ "dataset_trust_remote_code": false,
19
+ "context_size": null,
20
+ "normalize_activations": "none",
21
+ "training_condition": "mixed",
22
+ "training_notes": "Deception behavioral SAE \u2014 same-prompt behavioral sampling. Model: TinyLlama/TinyLlama-1.1B-intermediate-step-1431k-3T, Layer 12, jumprelu. See https://github.com/SolshineCode/deception-nanochat-sae-research"
23
+ }
tinyllama_jumprelu_L12_mixed/sae_weights.safetensors ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:6fe7a11876ae50ab489e8f577a68377f5f949c63937321578343232b1a9b7e56
3
+ size 134291856
tinyllama_jumprelu_L15_deceptive_only/cfg.json ADDED
@@ -0,0 +1,23 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "architecture": "jumprelu",
3
+ "d_in": 2048,
4
+ "d_sae": 8192,
5
+ "dtype": "float32",
6
+ "device": "cpu",
7
+ "model_name": "TinyLlama/TinyLlama-1.1B-intermediate-step-1431k-3T",
8
+ "hook_name": "model.layers.15",
9
+ "hook_layer": 15,
10
+ "hook_head_index": null,
11
+ "activation_fn_str": "jumprelu",
12
+ "activation_fn_kwargs": {},
13
+ "apply_b_dec_to_input": false,
14
+ "finetuning_scaling_factor": false,
15
+ "sae_lens_training_version": "deception-behavioral-v1",
16
+ "prepend_bos": false,
17
+ "dataset_path": "Solshine/deception-behavioral-multimodel",
18
+ "dataset_trust_remote_code": false,
19
+ "context_size": null,
20
+ "normalize_activations": "none",
21
+ "training_condition": "deceptive_only",
22
+ "training_notes": "Deception behavioral SAE \u2014 same-prompt behavioral sampling. Model: TinyLlama/TinyLlama-1.1B-intermediate-step-1431k-3T, Layer 15, jumprelu. See https://github.com/SolshineCode/deception-nanochat-sae-research"
23
+ }
tinyllama_jumprelu_L15_deceptive_only/sae_weights.safetensors ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:4823b3d400d80fc07d1ff1d22745981a8c49ce5efd083d4ed11f53c8eb3c8863
3
+ size 134291856
tinyllama_jumprelu_L15_honest_only/cfg.json ADDED
@@ -0,0 +1,23 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "architecture": "jumprelu",
3
+ "d_in": 2048,
4
+ "d_sae": 8192,
5
+ "dtype": "float32",
6
+ "device": "cpu",
7
+ "model_name": "TinyLlama/TinyLlama-1.1B-intermediate-step-1431k-3T",
8
+ "hook_name": "model.layers.15",
9
+ "hook_layer": 15,
10
+ "hook_head_index": null,
11
+ "activation_fn_str": "jumprelu",
12
+ "activation_fn_kwargs": {},
13
+ "apply_b_dec_to_input": false,
14
+ "finetuning_scaling_factor": false,
15
+ "sae_lens_training_version": "deception-behavioral-v1",
16
+ "prepend_bos": false,
17
+ "dataset_path": "Solshine/deception-behavioral-multimodel",
18
+ "dataset_trust_remote_code": false,
19
+ "context_size": null,
20
+ "normalize_activations": "none",
21
+ "training_condition": "honest_only",
22
+ "training_notes": "Deception behavioral SAE \u2014 same-prompt behavioral sampling. Model: TinyLlama/TinyLlama-1.1B-intermediate-step-1431k-3T, Layer 15, jumprelu. See https://github.com/SolshineCode/deception-nanochat-sae-research"
23
+ }
tinyllama_jumprelu_L15_honest_only/sae_weights.safetensors ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:e6f23c96daf192e923033e77d78df488e17f56e01ee15aeccfd1aacfc917941f
3
+ size 134291856
tinyllama_jumprelu_L15_mixed/cfg.json ADDED
@@ -0,0 +1,23 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "architecture": "jumprelu",
3
+ "d_in": 2048,
4
+ "d_sae": 8192,
5
+ "dtype": "float32",
6
+ "device": "cpu",
7
+ "model_name": "TinyLlama/TinyLlama-1.1B-intermediate-step-1431k-3T",
8
+ "hook_name": "model.layers.15",
9
+ "hook_layer": 15,
10
+ "hook_head_index": null,
11
+ "activation_fn_str": "jumprelu",
12
+ "activation_fn_kwargs": {},
13
+ "apply_b_dec_to_input": false,
14
+ "finetuning_scaling_factor": false,
15
+ "sae_lens_training_version": "deception-behavioral-v1",
16
+ "prepend_bos": false,
17
+ "dataset_path": "Solshine/deception-behavioral-multimodel",
18
+ "dataset_trust_remote_code": false,
19
+ "context_size": null,
20
+ "normalize_activations": "none",
21
+ "training_condition": "mixed",
22
+ "training_notes": "Deception behavioral SAE \u2014 same-prompt behavioral sampling. Model: TinyLlama/TinyLlama-1.1B-intermediate-step-1431k-3T, Layer 15, jumprelu. See https://github.com/SolshineCode/deception-nanochat-sae-research"
23
+ }
tinyllama_jumprelu_L15_mixed/sae_weights.safetensors ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:6630d4f2cb06eed8e0d9aee4ad95c55337e3bac0e089ad1d9110e43d36ab508e
3
+ size 134291856
tinyllama_jumprelu_L3_deceptive_only/cfg.json ADDED
@@ -0,0 +1,23 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "architecture": "jumprelu",
3
+ "d_in": 2048,
4
+ "d_sae": 8192,
5
+ "dtype": "float32",
6
+ "device": "cpu",
7
+ "model_name": "TinyLlama/TinyLlama-1.1B-intermediate-step-1431k-3T",
8
+ "hook_name": "model.layers.3",
9
+ "hook_layer": 3,
10
+ "hook_head_index": null,
11
+ "activation_fn_str": "jumprelu",
12
+ "activation_fn_kwargs": {},
13
+ "apply_b_dec_to_input": false,
14
+ "finetuning_scaling_factor": false,
15
+ "sae_lens_training_version": "deception-behavioral-v1",
16
+ "prepend_bos": false,
17
+ "dataset_path": "Solshine/deception-behavioral-multimodel",
18
+ "dataset_trust_remote_code": false,
19
+ "context_size": null,
20
+ "normalize_activations": "none",
21
+ "training_condition": "deceptive_only",
22
+ "training_notes": "Deception behavioral SAE \u2014 same-prompt behavioral sampling. Model: TinyLlama/TinyLlama-1.1B-intermediate-step-1431k-3T, Layer 3, jumprelu. See https://github.com/SolshineCode/deception-nanochat-sae-research"
23
+ }
tinyllama_jumprelu_L3_deceptive_only/sae_weights.safetensors ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:94d9e4d2310ffaed9a1e37a6257312ca23f6c0a70927073bf956a3c251efac4d
3
+ size 134291856
tinyllama_jumprelu_L3_honest_only/cfg.json ADDED
@@ -0,0 +1,23 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "architecture": "jumprelu",
3
+ "d_in": 2048,
4
+ "d_sae": 8192,
5
+ "dtype": "float32",
6
+ "device": "cpu",
7
+ "model_name": "TinyLlama/TinyLlama-1.1B-intermediate-step-1431k-3T",
8
+ "hook_name": "model.layers.3",
9
+ "hook_layer": 3,
10
+ "hook_head_index": null,
11
+ "activation_fn_str": "jumprelu",
12
+ "activation_fn_kwargs": {},
13
+ "apply_b_dec_to_input": false,
14
+ "finetuning_scaling_factor": false,
15
+ "sae_lens_training_version": "deception-behavioral-v1",
16
+ "prepend_bos": false,
17
+ "dataset_path": "Solshine/deception-behavioral-multimodel",
18
+ "dataset_trust_remote_code": false,
19
+ "context_size": null,
20
+ "normalize_activations": "none",
21
+ "training_condition": "honest_only",
22
+ "training_notes": "Deception behavioral SAE \u2014 same-prompt behavioral sampling. Model: TinyLlama/TinyLlama-1.1B-intermediate-step-1431k-3T, Layer 3, jumprelu. See https://github.com/SolshineCode/deception-nanochat-sae-research"
23
+ }
tinyllama_jumprelu_L3_honest_only/sae_weights.safetensors ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:9f8cbefe95c250dba27540d297af4f275813d6703a5fcf92697ffc91f5701af7
3
+ size 134291856
tinyllama_jumprelu_L3_mixed/cfg.json ADDED
@@ -0,0 +1,23 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "architecture": "jumprelu",
3
+ "d_in": 2048,
4
+ "d_sae": 8192,
5
+ "dtype": "float32",
6
+ "device": "cpu",
7
+ "model_name": "TinyLlama/TinyLlama-1.1B-intermediate-step-1431k-3T",
8
+ "hook_name": "model.layers.3",
9
+ "hook_layer": 3,
10
+ "hook_head_index": null,
11
+ "activation_fn_str": "jumprelu",
12
+ "activation_fn_kwargs": {},
13
+ "apply_b_dec_to_input": false,
14
+ "finetuning_scaling_factor": false,
15
+ "sae_lens_training_version": "deception-behavioral-v1",
16
+ "prepend_bos": false,
17
+ "dataset_path": "Solshine/deception-behavioral-multimodel",
18
+ "dataset_trust_remote_code": false,
19
+ "context_size": null,
20
+ "normalize_activations": "none",
21
+ "training_condition": "mixed",
22
+ "training_notes": "Deception behavioral SAE \u2014 same-prompt behavioral sampling. Model: TinyLlama/TinyLlama-1.1B-intermediate-step-1431k-3T, Layer 3, jumprelu. See https://github.com/SolshineCode/deception-nanochat-sae-research"
23
+ }
tinyllama_jumprelu_L3_mixed/sae_weights.safetensors ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:7a6783e9733e4ca00f38cb297f252dbe29ea6634c6b1d54b6383c39a2cc2ea99
3
+ size 134291856
tinyllama_jumprelu_L6_deceptive_only/cfg.json ADDED
@@ -0,0 +1,23 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "architecture": "jumprelu",
3
+ "d_in": 2048,
4
+ "d_sae": 8192,
5
+ "dtype": "float32",
6
+ "device": "cpu",
7
+ "model_name": "TinyLlama/TinyLlama-1.1B-intermediate-step-1431k-3T",
8
+ "hook_name": "model.layers.6",
9
+ "hook_layer": 6,
10
+ "hook_head_index": null,
11
+ "activation_fn_str": "jumprelu",
12
+ "activation_fn_kwargs": {},
13
+ "apply_b_dec_to_input": false,
14
+ "finetuning_scaling_factor": false,
15
+ "sae_lens_training_version": "deception-behavioral-v1",
16
+ "prepend_bos": false,
17
+ "dataset_path": "Solshine/deception-behavioral-multimodel",
18
+ "dataset_trust_remote_code": false,
19
+ "context_size": null,
20
+ "normalize_activations": "none",
21
+ "training_condition": "deceptive_only",
22
+ "training_notes": "Deception behavioral SAE \u2014 same-prompt behavioral sampling. Model: TinyLlama/TinyLlama-1.1B-intermediate-step-1431k-3T, Layer 6, jumprelu. See https://github.com/SolshineCode/deception-nanochat-sae-research"
23
+ }
tinyllama_jumprelu_L6_deceptive_only/sae_weights.safetensors ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:45262e4633200f29bd6f59ed2dd44f770fe23994af82d7cfb0d63e520701a18f
3
+ size 134291856
tinyllama_jumprelu_L6_honest_only/cfg.json ADDED
@@ -0,0 +1,23 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "architecture": "jumprelu",
3
+ "d_in": 2048,
4
+ "d_sae": 8192,
5
+ "dtype": "float32",
6
+ "device": "cpu",
7
+ "model_name": "TinyLlama/TinyLlama-1.1B-intermediate-step-1431k-3T",
8
+ "hook_name": "model.layers.6",
9
+ "hook_layer": 6,
10
+ "hook_head_index": null,
11
+ "activation_fn_str": "jumprelu",
12
+ "activation_fn_kwargs": {},
13
+ "apply_b_dec_to_input": false,
14
+ "finetuning_scaling_factor": false,
15
+ "sae_lens_training_version": "deception-behavioral-v1",
16
+ "prepend_bos": false,
17
+ "dataset_path": "Solshine/deception-behavioral-multimodel",
18
+ "dataset_trust_remote_code": false,
19
+ "context_size": null,
20
+ "normalize_activations": "none",
21
+ "training_condition": "honest_only",
22
+ "training_notes": "Deception behavioral SAE \u2014 same-prompt behavioral sampling. Model: TinyLlama/TinyLlama-1.1B-intermediate-step-1431k-3T, Layer 6, jumprelu. See https://github.com/SolshineCode/deception-nanochat-sae-research"
23
+ }
tinyllama_jumprelu_L6_honest_only/sae_weights.safetensors ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:6a4b54adc611208dac0e901ea6506574a9725d4eb935390c5ae49f183f912be3
3
+ size 134291856
tinyllama_jumprelu_L6_mixed/cfg.json ADDED
@@ -0,0 +1,23 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "architecture": "jumprelu",
3
+ "d_in": 2048,
4
+ "d_sae": 8192,
5
+ "dtype": "float32",
6
+ "device": "cpu",
7
+ "model_name": "TinyLlama/TinyLlama-1.1B-intermediate-step-1431k-3T",
8
+ "hook_name": "model.layers.6",
9
+ "hook_layer": 6,
10
+ "hook_head_index": null,
11
+ "activation_fn_str": "jumprelu",
12
+ "activation_fn_kwargs": {},
13
+ "apply_b_dec_to_input": false,
14
+ "finetuning_scaling_factor": false,
15
+ "sae_lens_training_version": "deception-behavioral-v1",
16
+ "prepend_bos": false,
17
+ "dataset_path": "Solshine/deception-behavioral-multimodel",
18
+ "dataset_trust_remote_code": false,
19
+ "context_size": null,
20
+ "normalize_activations": "none",
21
+ "training_condition": "mixed",
22
+ "training_notes": "Deception behavioral SAE \u2014 same-prompt behavioral sampling. Model: TinyLlama/TinyLlama-1.1B-intermediate-step-1431k-3T, Layer 6, jumprelu. See https://github.com/SolshineCode/deception-nanochat-sae-research"
23
+ }
tinyllama_jumprelu_L6_mixed/sae_weights.safetensors ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:8fbc85a9a08b1a6c843e60d2850cfbb45cd7121455a9449e483902da3ffdfc79
3
+ size 134291856
tinyllama_jumprelu_L9_deceptive_only/cfg.json ADDED
@@ -0,0 +1,23 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "architecture": "jumprelu",
3
+ "d_in": 2048,
4
+ "d_sae": 8192,
5
+ "dtype": "float32",
6
+ "device": "cpu",
7
+ "model_name": "TinyLlama/TinyLlama-1.1B-intermediate-step-1431k-3T",
8
+ "hook_name": "model.layers.9",
9
+ "hook_layer": 9,
10
+ "hook_head_index": null,
11
+ "activation_fn_str": "jumprelu",
12
+ "activation_fn_kwargs": {},
13
+ "apply_b_dec_to_input": false,
14
+ "finetuning_scaling_factor": false,
15
+ "sae_lens_training_version": "deception-behavioral-v1",
16
+ "prepend_bos": false,
17
+ "dataset_path": "Solshine/deception-behavioral-multimodel",
18
+ "dataset_trust_remote_code": false,
19
+ "context_size": null,
20
+ "normalize_activations": "none",
21
+ "training_condition": "deceptive_only",
22
+ "training_notes": "Deception behavioral SAE \u2014 same-prompt behavioral sampling. Model: TinyLlama/TinyLlama-1.1B-intermediate-step-1431k-3T, Layer 9, jumprelu. See https://github.com/SolshineCode/deception-nanochat-sae-research"
23
+ }
tinyllama_jumprelu_L9_deceptive_only/sae_weights.safetensors ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:d5a2be2fed82c51089d1a2559fb1658c73e865531d3f243f2f58b85edded3086
3
+ size 134291856
tinyllama_jumprelu_L9_honest_only/cfg.json ADDED
@@ -0,0 +1,23 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "architecture": "jumprelu",
3
+ "d_in": 2048,
4
+ "d_sae": 8192,
5
+ "dtype": "float32",
6
+ "device": "cpu",
7
+ "model_name": "TinyLlama/TinyLlama-1.1B-intermediate-step-1431k-3T",
8
+ "hook_name": "model.layers.9",
9
+ "hook_layer": 9,
10
+ "hook_head_index": null,
11
+ "activation_fn_str": "jumprelu",
12
+ "activation_fn_kwargs": {},
13
+ "apply_b_dec_to_input": false,
14
+ "finetuning_scaling_factor": false,
15
+ "sae_lens_training_version": "deception-behavioral-v1",
16
+ "prepend_bos": false,
17
+ "dataset_path": "Solshine/deception-behavioral-multimodel",
18
+ "dataset_trust_remote_code": false,
19
+ "context_size": null,
20
+ "normalize_activations": "none",
21
+ "training_condition": "honest_only",
22
+ "training_notes": "Deception behavioral SAE \u2014 same-prompt behavioral sampling. Model: TinyLlama/TinyLlama-1.1B-intermediate-step-1431k-3T, Layer 9, jumprelu. See https://github.com/SolshineCode/deception-nanochat-sae-research"
23
+ }
tinyllama_jumprelu_L9_honest_only/sae_weights.safetensors ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:df4739d48a704e527d6aba02f74b4e7c96947b2422ac99cda5388e2a848ff0f4
3
+ size 134291856
tinyllama_jumprelu_L9_mixed/cfg.json ADDED
@@ -0,0 +1,23 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "architecture": "jumprelu",
3
+ "d_in": 2048,
4
+ "d_sae": 8192,
5
+ "dtype": "float32",
6
+ "device": "cpu",
7
+ "model_name": "TinyLlama/TinyLlama-1.1B-intermediate-step-1431k-3T",
8
+ "hook_name": "model.layers.9",
9
+ "hook_layer": 9,
10
+ "hook_head_index": null,
11
+ "activation_fn_str": "jumprelu",
12
+ "activation_fn_kwargs": {},
13
+ "apply_b_dec_to_input": false,
14
+ "finetuning_scaling_factor": false,
15
+ "sae_lens_training_version": "deception-behavioral-v1",
16
+ "prepend_bos": false,
17
+ "dataset_path": "Solshine/deception-behavioral-multimodel",
18
+ "dataset_trust_remote_code": false,
19
+ "context_size": null,
20
+ "normalize_activations": "none",
21
+ "training_condition": "mixed",
22
+ "training_notes": "Deception behavioral SAE \u2014 same-prompt behavioral sampling. Model: TinyLlama/TinyLlama-1.1B-intermediate-step-1431k-3T, Layer 9, jumprelu. See https://github.com/SolshineCode/deception-nanochat-sae-research"
23
+ }
tinyllama_jumprelu_L9_mixed/sae_weights.safetensors ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:fb2e38fa340ff9bd12eddb52d3e8cb747df912a7e367df386ce973c2e2893ac4
3
+ size 134291856
tinyllama_jumprelu_ste_L12_deceptive_only/cfg.json ADDED
@@ -0,0 +1,23 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "architecture": "jumprelu",
3
+ "d_in": 2048,
4
+ "d_sae": 8192,
5
+ "dtype": "float32",
6
+ "device": "cpu",
7
+ "model_name": "TinyLlama/TinyLlama-1.1B-intermediate-step-1431k-3T",
8
+ "hook_name": "model.layers.12",
9
+ "hook_layer": 12,
10
+ "hook_head_index": null,
11
+ "activation_fn_str": "jumprelu",
12
+ "activation_fn_kwargs": {},
13
+ "apply_b_dec_to_input": false,
14
+ "finetuning_scaling_factor": false,
15
+ "sae_lens_training_version": "deception-behavioral-v1",
16
+ "prepend_bos": false,
17
+ "dataset_path": "Solshine/deception-behavioral-multimodel",
18
+ "dataset_trust_remote_code": false,
19
+ "context_size": null,
20
+ "normalize_activations": "none",
21
+ "training_condition": "deceptive_only",
22
+ "training_notes": "STE validation SAE (2026-04-11) \u2014 deceptive_only training, Gaussian-kernel STE fix (arXiv:2407.14435). See https://github.com/SolshineCode/deception-nanochat-sae-research"
23
+ }
tinyllama_jumprelu_ste_L12_deceptive_only/sae_weights.safetensors ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:6ee617b4a3060449d7fe414c34d40aa82eed70a7c3cedd1ccf0dd0e876f69cf0
3
+ size 134291856
tinyllama_jumprelu_ste_L12_honest_only/cfg.json ADDED
@@ -0,0 +1,23 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "architecture": "jumprelu",
3
+ "d_in": 2048,
4
+ "d_sae": 8192,
5
+ "dtype": "float32",
6
+ "device": "cpu",
7
+ "model_name": "TinyLlama/TinyLlama-1.1B-intermediate-step-1431k-3T",
8
+ "hook_name": "model.layers.12",
9
+ "hook_layer": 12,
10
+ "hook_head_index": null,
11
+ "activation_fn_str": "jumprelu",
12
+ "activation_fn_kwargs": {},
13
+ "apply_b_dec_to_input": false,
14
+ "finetuning_scaling_factor": false,
15
+ "sae_lens_training_version": "deception-behavioral-v1",
16
+ "prepend_bos": false,
17
+ "dataset_path": "Solshine/deception-behavioral-multimodel",
18
+ "dataset_trust_remote_code": false,
19
+ "context_size": null,
20
+ "normalize_activations": "none",
21
+ "training_condition": "honest_only",
22
+ "training_notes": "STE validation SAE (2026-04-11) \u2014 honest_only training, Gaussian-kernel STE fix (arXiv:2407.14435). See https://github.com/SolshineCode/deception-nanochat-sae-research"
23
+ }
tinyllama_jumprelu_ste_L12_honest_only/sae_weights.safetensors ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:c14e5fd731c2a3d1796f319fe27f5a4da56b077b9492afc1e10aceed0a4497e0
3
+ size 134291856
tinyllama_jumprelu_ste_L12_mixed/cfg.json ADDED
@@ -0,0 +1,23 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "architecture": "jumprelu",
3
+ "d_in": 2048,
4
+ "d_sae": 8192,
5
+ "dtype": "float32",
6
+ "device": "cpu",
7
+ "model_name": "TinyLlama/TinyLlama-1.1B-intermediate-step-1431k-3T",
8
+ "hook_name": "model.layers.12",
9
+ "hook_layer": 12,
10
+ "hook_head_index": null,
11
+ "activation_fn_str": "jumprelu",
12
+ "activation_fn_kwargs": {},
13
+ "apply_b_dec_to_input": false,
14
+ "finetuning_scaling_factor": false,
15
+ "sae_lens_training_version": "deception-behavioral-v1",
16
+ "prepend_bos": false,
17
+ "dataset_path": "Solshine/deception-behavioral-multimodel",
18
+ "dataset_trust_remote_code": false,
19
+ "context_size": null,
20
+ "normalize_activations": "none",
21
+ "training_condition": "mixed",
22
+ "training_notes": "STE validation SAE (2026-04-11) \u2014 mixed training, Gaussian-kernel STE fix (arXiv:2407.14435). See https://github.com/SolshineCode/deception-nanochat-sae-research"
23
+ }
tinyllama_jumprelu_ste_L12_mixed/sae_weights.safetensors ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:914010cbb4f201069d940a09e69405804a14742ff22275cbd0489b42e1ec5a90
3
+ size 134291856
tinyllama_jumprelu_ste_L15_deceptive_only/cfg.json ADDED
@@ -0,0 +1,23 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "architecture": "jumprelu",
3
+ "d_in": 2048,
4
+ "d_sae": 8192,
5
+ "dtype": "float32",
6
+ "device": "cpu",
7
+ "model_name": "TinyLlama/TinyLlama-1.1B-intermediate-step-1431k-3T",
8
+ "hook_name": "model.layers.15",
9
+ "hook_layer": 15,
10
+ "hook_head_index": null,
11
+ "activation_fn_str": "jumprelu",
12
+ "activation_fn_kwargs": {},
13
+ "apply_b_dec_to_input": false,
14
+ "finetuning_scaling_factor": false,
15
+ "sae_lens_training_version": "deception-behavioral-v1",
16
+ "prepend_bos": false,
17
+ "dataset_path": "Solshine/deception-behavioral-multimodel",
18
+ "dataset_trust_remote_code": false,
19
+ "context_size": null,
20
+ "normalize_activations": "none",
21
+ "training_condition": "deceptive_only",
22
+ "training_notes": "STE validation SAE (2026-04-11) \u2014 deceptive_only training, Gaussian-kernel STE fix (arXiv:2407.14435). See https://github.com/SolshineCode/deception-nanochat-sae-research"
23
+ }
tinyllama_jumprelu_ste_L15_deceptive_only/sae_weights.safetensors ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:0875860bc89d82807afaf97c975c4adba8726752f4886091ccc9117d1e89a37f
3
+ size 134291856
tinyllama_jumprelu_ste_L15_honest_only/cfg.json ADDED
@@ -0,0 +1,23 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "architecture": "jumprelu",
3
+ "d_in": 2048,
4
+ "d_sae": 8192,
5
+ "dtype": "float32",
6
+ "device": "cpu",
7
+ "model_name": "TinyLlama/TinyLlama-1.1B-intermediate-step-1431k-3T",
8
+ "hook_name": "model.layers.15",
9
+ "hook_layer": 15,
10
+ "hook_head_index": null,
11
+ "activation_fn_str": "jumprelu",
12
+ "activation_fn_kwargs": {},
13
+ "apply_b_dec_to_input": false,
14
+ "finetuning_scaling_factor": false,
15
+ "sae_lens_training_version": "deception-behavioral-v1",
16
+ "prepend_bos": false,
17
+ "dataset_path": "Solshine/deception-behavioral-multimodel",
18
+ "dataset_trust_remote_code": false,
19
+ "context_size": null,
20
+ "normalize_activations": "none",
21
+ "training_condition": "honest_only",
22
+ "training_notes": "STE validation SAE (2026-04-11) \u2014 honest_only training, Gaussian-kernel STE fix (arXiv:2407.14435). See https://github.com/SolshineCode/deception-nanochat-sae-research"
23
+ }
tinyllama_jumprelu_ste_L15_honest_only/sae_weights.safetensors ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:a5794479e77cec1e6063f32fefcb34b0678a443c06eab638df1733a0613ff684
3
+ size 134291856
tinyllama_jumprelu_ste_L15_mixed/cfg.json ADDED
@@ -0,0 +1,23 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "architecture": "jumprelu",
3
+ "d_in": 2048,
4
+ "d_sae": 8192,
5
+ "dtype": "float32",
6
+ "device": "cpu",
7
+ "model_name": "TinyLlama/TinyLlama-1.1B-intermediate-step-1431k-3T",
8
+ "hook_name": "model.layers.15",
9
+ "hook_layer": 15,
10
+ "hook_head_index": null,
11
+ "activation_fn_str": "jumprelu",
12
+ "activation_fn_kwargs": {},
13
+ "apply_b_dec_to_input": false,
14
+ "finetuning_scaling_factor": false,
15
+ "sae_lens_training_version": "deception-behavioral-v1",
16
+ "prepend_bos": false,
17
+ "dataset_path": "Solshine/deception-behavioral-multimodel",
18
+ "dataset_trust_remote_code": false,
19
+ "context_size": null,
20
+ "normalize_activations": "none",
21
+ "training_condition": "mixed",
22
+ "training_notes": "STE validation SAE (2026-04-11) \u2014 mixed training, Gaussian-kernel STE fix (arXiv:2407.14435). See https://github.com/SolshineCode/deception-nanochat-sae-research"
23
+ }
tinyllama_jumprelu_ste_L15_mixed/sae_weights.safetensors ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:7bd22cac426c5177990b3891bc0523a8a44077a4e7f709796f7f530a6d06b449
3
+ size 134291856
tinyllama_jumprelu_ste_L9_deceptive_only/cfg.json ADDED
@@ -0,0 +1,23 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "architecture": "jumprelu",
3
+ "d_in": 2048,
4
+ "d_sae": 8192,
5
+ "dtype": "float32",
6
+ "device": "cpu",
7
+ "model_name": "TinyLlama/TinyLlama-1.1B-intermediate-step-1431k-3T",
8
+ "hook_name": "model.layers.9",
9
+ "hook_layer": 9,
10
+ "hook_head_index": null,
11
+ "activation_fn_str": "jumprelu",
12
+ "activation_fn_kwargs": {},
13
+ "apply_b_dec_to_input": false,
14
+ "finetuning_scaling_factor": false,
15
+ "sae_lens_training_version": "deception-behavioral-v1",
16
+ "prepend_bos": false,
17
+ "dataset_path": "Solshine/deception-behavioral-multimodel",
18
+ "dataset_trust_remote_code": false,
19
+ "context_size": null,
20
+ "normalize_activations": "none",
21
+ "training_condition": "deceptive_only",
22
+ "training_notes": "STE validation SAE (2026-04-11) \u2014 deceptive_only training, Gaussian-kernel STE fix (arXiv:2407.14435). See https://github.com/SolshineCode/deception-nanochat-sae-research"
23
+ }
tinyllama_jumprelu_ste_L9_deceptive_only/sae_weights.safetensors ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:80805a3416a594471b6ff3669fb14c21fb115080ef59b1d2ed776f90d35aa9e6
3
+ size 134291856
tinyllama_jumprelu_ste_L9_honest_only/cfg.json ADDED
@@ -0,0 +1,23 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "architecture": "jumprelu",
3
+ "d_in": 2048,
4
+ "d_sae": 8192,
5
+ "dtype": "float32",
6
+ "device": "cpu",
7
+ "model_name": "TinyLlama/TinyLlama-1.1B-intermediate-step-1431k-3T",
8
+ "hook_name": "model.layers.9",
9
+ "hook_layer": 9,
10
+ "hook_head_index": null,
11
+ "activation_fn_str": "jumprelu",
12
+ "activation_fn_kwargs": {},
13
+ "apply_b_dec_to_input": false,
14
+ "finetuning_scaling_factor": false,
15
+ "sae_lens_training_version": "deception-behavioral-v1",
16
+ "prepend_bos": false,
17
+ "dataset_path": "Solshine/deception-behavioral-multimodel",
18
+ "dataset_trust_remote_code": false,
19
+ "context_size": null,
20
+ "normalize_activations": "none",
21
+ "training_condition": "honest_only",
22
+ "training_notes": "STE validation SAE (2026-04-11) \u2014 honest_only training, Gaussian-kernel STE fix (arXiv:2407.14435). See https://github.com/SolshineCode/deception-nanochat-sae-research"
23
+ }
tinyllama_jumprelu_ste_L9_honest_only/sae_weights.safetensors ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:5d5b2a60730506899627e5a980ff871cc0d5d98984de4025d2f0b6a7147ae5bd
3
+ size 134291856
tinyllama_jumprelu_ste_L9_mixed/cfg.json ADDED
@@ -0,0 +1,23 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "architecture": "jumprelu",
3
+ "d_in": 2048,
4
+ "d_sae": 8192,
5
+ "dtype": "float32",
6
+ "device": "cpu",
7
+ "model_name": "TinyLlama/TinyLlama-1.1B-intermediate-step-1431k-3T",
8
+ "hook_name": "model.layers.9",
9
+ "hook_layer": 9,
10
+ "hook_head_index": null,
11
+ "activation_fn_str": "jumprelu",
12
+ "activation_fn_kwargs": {},
13
+ "apply_b_dec_to_input": false,
14
+ "finetuning_scaling_factor": false,
15
+ "sae_lens_training_version": "deception-behavioral-v1",
16
+ "prepend_bos": false,
17
+ "dataset_path": "Solshine/deception-behavioral-multimodel",
18
+ "dataset_trust_remote_code": false,
19
+ "context_size": null,
20
+ "normalize_activations": "none",
21
+ "training_condition": "mixed",
22
+ "training_notes": "STE validation SAE (2026-04-11) \u2014 mixed training, Gaussian-kernel STE fix (arXiv:2407.14435). See https://github.com/SolshineCode/deception-nanochat-sae-research"
23
+ }
tinyllama_jumprelu_ste_L9_mixed/sae_weights.safetensors ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:e328b9c3d999025a060236c8492f075a5e35a0f5900940adf058dc3b87cf5cd7
3
+ size 134291856