Qwen3Loop-0.6B-Tool-Calling-SFT (E013_pareto_bundle)

Qwen3Loop-0.6B-Tool-Calling-SFT has been updated from the earlier v3 SFT weights to E013_pareto_bundle, an unrolled 56-layer standard Qwen3ForCausalLM checkpoint combining:

  1. SFT + L2 RL (Step 61) trained on multi-step tool calling, abstention, and mid-trajectory recovery.
  2. Surgical Layer-49 zero-hole attention repair (E001), removing the late-pass collapse present in the raw RL checkpoint.
  3. Orthogonal Pareto-KKT lm_head weight-space bundles (E002โ€“E013), touching 6,144 / 1,192,098,816 (0.000515%) parameters across six output rows (19, 1703, 12509, 17453, 23657, 46254).

Note on llama.cpp -hf tags: All GGUF filenames follow the GGUF naming convention with a hyphen before the quant tag (e.g., -BF16.gguf, -Q8_0.gguf, -Q4_K_M.gguf). Because llama.cpp -hf without an explicit :TAG often defaults to :Q4_K_Mโ€”and 4-bit quantization degrades safety and reliability at this parameter countโ€”always pass an explicit tag such as :BF16 or :F32 (or :Q8_0 under tighter VRAM limits).


Architecture

  • Hugging Face architecture: standard unrolled Qwen3ForCausalLM (model_type: qwen3, natively supported by stock transformers, llama.cpp, vLLM, Ollama, and LM Studio with no trust_remote_code or runtime patches).
  • Recurrent origin: 28 physical blocks unrolled across the recurrent middle stack (7 prefix + 14 ร— 3 loop passes + 7 suffix = 56 logical layers, hidden_size: 1024, intermediate_size: 3072, 16 attention heads, 8 KV heads, max_position_embeddings: 40960).
  • Untied output head: tie_word_embeddings: false because E002โ€“E013 optimize 43 lm_head.weight rows independently of model.embed_tokens.weight. The recurrent triple-loop middle stack (layers 7..20, 21..34, 35..48) remains 100% weight-identical (max_diff = 0.0), while E001 repairs suffix layer 49 (physical layer 21, outside the loop).
  • Earlier v3 release: The older 28-block folded SFT checkpoint (Qwen3LoopForCausalLM, engine/ patches, latent_halting_probe.pt) remains preserved in git history.

Measured benchmark (tool-eval-bench, seed 42, 69 scenarios, 16k context)

Evaluated locally on RTX 3060 via llama-server (--jinja -ngl 999 -c 16384 --flash-attn on):

Artifact / Precision File Size Score Points Pass / Partial / Fail Safety Warnings Recommendation
Step 61 baseline (E000, F32) โ€” 24% 33/136 โ€” 3 Superseded baseline
E013 F32 (-F32.gguf) 4.77 GB 66% 90/136 35 / 20 / 14 0 Canonical reference precision
E013 BF16 (-BF16.gguf) 2.39 GB 65% 88/136 34 / 20 / 15 0 Recommended half-precision default
E013 F16 (-F16.gguf) 2.39 GB 64% 87/136 34 / 19 / 16 0 Wide FP16 backend compatibility
E013 Q8_0 (-Q8_0.gguf, F32 output.weight) 1.73 GB 59% 80/136 32 / 16 / 21 0 Smaller footprint (output.weight kept in F32; vanilla Q8_0 drops to 57% / 77/136)
E013 Q4_K_M (-Q4_K_M.gguf) 0.75 GB 48% 66/138 23 / 20 / 26 2 Not recommended for agentic use (TC-48, TC-49 safety warnings); kept for :Q4_K_M tag compatibility

Canonical E013 F32 (FF4953BBAD048859405EFE87B40D57EE60DA60080E5C91DDB18CD6F83FE79E5E) breakdown:

  • deployability: 73, responsiveness: 89, 0 regressions relative to E011 (80/136) and E012 (82/136).
  • Per-category points: A=6/6, B=4/6, C=3/8, D=6/6, E=2/6, F=4/6, G=4/6, H=7/8, I=10/20, J=5/6, K=17/26, L=6/8, M=3/6, N=4/6, O=9/12.

Explicit limitations

  1. Quantization sensitivity: Because E002โ€“E013 rely on small float32 lm_head.weight logit margin edits (0.0168โ€“0.2710 L2 norm per edited row), 4-bit body quantization (:Q4_K_M) shifts hidden-state geometry enough to trigger early-send safety warnings (TC-48, TC-49). Use :F32 or :BF16 whenever memory allows.
  2. Remaining 0.6B failure modes: On seed 42 F32, 14 of 69 scenarios still fail and 20 receive partial creditโ€”mostly complex argument extraction, multi-turn planning depth, and unfamiliar tool schemas.
  3. English primary: Training and tool-calling validation are overwhelmingly in English; non-English prompts degrade tool reliability.
  4. External gating advised for irreversible actions: Pair the model with application-level confirmation checks before destructive writes, payments, or outbound emails.

Files

GGUF (llama.cpp, Ollama, LM Studio)

Hugging Face Safetensors (transformers / vLLM)


Quickstart

llama.cpp (-hf tag matching enabled)

llama-server -hf Lordnyx/qwen3loop-0.6b-tool-calling-sft-v1:BF16 --jinja -ngl 999 -c 16384 --flash-attn on

Other tags: :F32, :F16, :Q8_0 (and :Q4_K_M, not recommended for safety-sensitive agents).

PyTorch / transformers

import torch
from transformers import AutoModelForCausalLM, AutoTokenizer

repo_id = "Lordnyx/qwen3loop-0.6b-tool-calling-sft-v1"
tokenizer = AutoTokenizer.from_pretrained(repo_id)
model = AutoModelForCausalLM.from_pretrained(
    repo_id,
    torch_dtype=torch.bfloat16,
    device_map="auto",
)

tools = [
    {
        "type": "function",
        "function": {
            "name": "convert_measurement_units",
            "description": "Convert physical quantities between units.",
            "parameters": {
                "type": "object",
                "properties": {
                    "value": {"type": "number"},
                    "from_unit": {"type": "string"},
                    "to_unit": {"type": "string"},
                },
                "required": ["value", "from_unit", "to_unit"],
            },
        },
    }
]

prompt = tokenizer.apply_chat_template(
    [{"role": "user", "content": "Convert 42 km to miles."}],
    tools=tools,
    tokenize=False,
    add_generation_prompt=True,
)
inputs = tokenizer(prompt, return_tensors="pt").to(model.device)
out = model.generate(**inputs, max_new_tokens=512, do_sample=False)
print(tokenizer.decode(out[0][inputs["input_ids"].shape[1]:], skip_special_tokens=False))

(Note: generation_config.json preserves Qwen3's default sampling parameters temperature=0.6, top_p=0.95, top_k=20, while the tool-eval-bench scores above were measured with OpenAI-compatible --jinja server tool calling.)

Downloads last month
1,040
Safetensors
Model size
1B params
Tensor type
F32
ยท
Inference Providers NEW
This model isn't deployed by any Inference Provider. ๐Ÿ™‹ Ask for provider support

Model tree for Lordnyx/qwen3loop-0.6b-tool-calling-sft-v1

Quantizations
1 model

Collection including Lordnyx/qwen3loop-0.6b-tool-calling-sft-v1