Update README.md

README.md

@@ -1,47 +1,26 @@
 ---
 license: mit
 task_categories:
-  - visual-question-answering
-  - image-text-to-text
+  - text-generation
 language:
   - en
 size_categories:
-  - 10K<n<100K
+  - 1K<n<10K
 tags:
-  - multimodal
-  - vision-language
+  - agentic
+  - code
+  - software-engineering
   - reinforcement-learning
   - context-aware
   - contrastive
 ---
-# ContextRL-Multimodal-Qwen3-VL
-The multimodal training set for **ContextRL**, used to train
-**[ContextRL-Qwen3-VL-8B](https://huggingface.co/xupy21/ContextRL_Qwen3_VL_8B)**, from
-the paper **Context-Aware RL for Agentic and Multimodal LLMs**. It is formatted for the
-Qwen3-VL chat template (with thinking format).
-## Composition
-The set contains **45,000 examples** combining two objectives:
-- **38,000 standard single-image QA examples** — used for the GRPO objective (`L_GRPO`).
-- **7,000 contrastive image pairs** — used for the context-selection objective (`L_CA`).
-  Each pair shares a (question, answer) reference but presents an anchor image and a
-  minimally-perturbed confounder image that supports a different answer. Pairs are built
-  via **generative editing** (natural images) and **similarity-based retrieval**
-  (structured/text-rich images).
+# ContextRL-Agentic
+The agentic (long-horizon) training set for **ContextRL**, used to train
+**[ContextRL-Klear-AgentForge-8B](https://huggingface.co/xupy21/ContextRL_Klear_AgentForge_8B)**
+and **[ContextRL-Qwen3-8B-Agentic](https://huggingface.co/xupy21/ContextRL_Qwen3_8B_Agentic)**,
+from the paper **Context-Aware RL for Agentic and Multimodal LLMs**.
 
-The data spans five visual subgroups:
-
-| Subgroup | Count | % |
-|---|---|---|
-| Geometry | 20,729 | 46.06 |
-| Non-geometric Math | 7,733 | 17.18 |
-| Natural Image | 5,930 | 13.18 |
-| Chart | 5,525 | 12.28 |
-| Science | 5,083 | 11.30 |
-| **Total** | **45,000** | **100.00** |
-
-Source datasets include MAVIS-Geometry, Geo170K, R-CoT, Geometry3K, GeoQA3, GeomVerse,
-MAVIS-Function, ICON-QA, MMK12, DVQA, ChartQA, M3CoT, ScienceQA, Visual-CoT, and ViRL39K.
 ## Setup
 Training and evaluation code, data construction pipelines, and detailed configurations are
 available in the repository:
-👉 **https://github.com/xupy2003/ContextAwareRL**
+👉 **https://github.com/xupy2003/ContextAwareRL**