Update README.md

README.md

@@ -1,69 +1,68 @@
+---
+license: mit
+base_model:
+- Qwen/Qwen3-4B
+pipeline_tag: text-classification
+---
+
 # OpenHands Critic 4B v1.0
 
-A 4B parameter critic model for evaluating AI agent trajectories, trained to predict task success from behavioral rubrics.
+A 4B parameter critic model for evaluating AI agent trajectories, trained to predict behavioral rubrics and task success.
+
+## Related Links
+
+- Paper: https://arxiv.org/abs/2603.03800
+- Rubrics (definitions & prompts): https://github.com/OpenHands/critic-rubrics
+- Docs (Use it in OpenHands Software Agent SDK): https://docs.openhands.dev/sdk/guides/critic
+- Docs (Use it in OpenHands CLI): https://docs.openhands.dev/openhands/usage/cli/critic
 
 ## Model Details
 
-- **Base Model**: Qwen3-4B
-- **Training**: Full parameter fine-tuning with BCE loss
+- **Base Model**: Qwen/Qwen3-4B
+- **Training**: Full-parameter fine-tuning with BCE loss
 - **Context Length**: Trained on 64K, supports up to 256K tokens
 - **Task**: Multi-label classification (26 labels: 25 rubric features + 1 success prediction)
 
-## Paper
-
-This model is described in the paper: **"Rubric-Supervised Critics for Sparse Agent Feedback"**
-
-### Key Results (Mixed-Outcome Subset)
-- **+15.9 points** Best@8 improvement over random selection (73.8% vs 57.9%)
-- **0.83 MRR** - correct trajectory typically ranked first
-- **83% compute reduction** via adaptive rollout (1.36 attempts vs 8)
+## Serving (vLLM Classification API)
+
+We serve this model using vLLM’s classification task:
+
+```bash
+vllm serve <MODEL_PATH> \
+  --host 0.0.0.0 \
+  --port 8000 \
+  --api-key <API_KEY> \
+  --served-model-name <MODEL_NAME> \
+  --task classify \
+  --max-model-len 262144 \
+  --dtype bfloat16 \
+  --trust-remote-code \
+  --enable-prefix-caching
+```
 
 ## Usage
 
-This model is designed for use with vLLM's classification API:
-
-```python
-from openai import OpenAI
+We recommend using the **OpenHands SDK** for inference instead of calling the vLLM classification endpoint directly.
 
-client = OpenAI(
-    base_url="YOUR_VLLM_SERVER_URL/v1",
-    api_key="YOUR_API_KEY"
-)
+Follow the SDK guide: https://docs.openhands.dev/sdk/guides/critic
 
-# Format your trajectory as a conversation
-messages = [
-    {"role": "system", "content": "You are evaluating an AI agent's task attempt..."},
-    {"role": "user", "content": "Task: ..."},
-    {"role": "assistant", "content": "Agent actions..."}
-]
+In particular, reuse the SDK client implementation here (it already handles formatting and API calls): https://github.com/OpenHands/software-agent-sdk/blob/main/openhands-sdk/openhands/sdk/critic/impl/api/critic.py
 
-# Get classification scores
-response = client.classifications.create(
-    model="openhands-critic-4b-v1.0",
-    messages=messages
-)
-
-# The model outputs probabilities for 26 labels:
-# - Labels 0-24: Rubric features (behavioral indicators)
-# - Label 25: Success prediction (primary output for ranking)
-```
-
-## Training Data
-
-Trained on 154K segments from:
-- Production agent conversations (150K segments)
-- SWE-Gym benchmark trajectories (4K segments)
-
-## License
-
-Please refer to the Qwen3 license for base model terms.
+At a high level, you will:
+1. Start a critic server (see **Serving** section above)
+2. Configure the SDK to point to your critic endpoint + API key
+3. Call the SDK critic to score trajectories (returns rubric probabilities + success score)
 
 ## Citation
 
 ```bibtex
-@article{openhands2025critic,
-  title={Rubric-Supervised Critics for Sparse Agent Feedback},
-  author={OpenHands Team},
-  year={2025}
+@misc{wang2026rubricsupervisedcriticsparserealworld,
+      title={A Rubric-Supervised Critic from Sparse Real-World Outcomes}, 
+      author={Xingyao Wang and Valerie Chen and Heng Ji and Graham Neubig},
+      year={2026},
+      eprint={2603.03800},
+      archivePrefix={arXiv},
+      primaryClass={cs.AI},
+      url={https://arxiv.org/abs/2603.03800}, 
 }
-```
+```