Keep model card concise

README.md

@@ -16,26 +16,25 @@ tags:
 
 # SkillScreen-2B-SBR
 
-SkillScreen scores malicious agent skills directly from `SKILL.md` in one forward pass. This checkpoint uses the Source-Balanced Random (SBR) training split and an 8,192-token input budget.
+2B decision model for detecting malicious agent skills from `SKILL.md`. Trained on MaliciousSkillBench using Source-Balanced Random (SBR), with an 8K input budget.
 
-Fine-tuned from [OpenJev 2B](https://huggingface.co/AlexWortega/openjev/tree/a20448012c213128955ca0c693e7c943865cab77/qwen3.5-2b-nli-v5) with LoRA and a trained classification head. Seed 42, one epoch, checkpoint 426. Full merged weights are at the root. The original adapter and trained head are in `adapter/`.
+Based on [OpenJev 2B](https://huggingface.co/AlexWortega/openjev). Full weights are at the root. The original LoRA adapter and classification head are in `adapter/`.
 
 ## Use
 
-CUDA with BF16 support.
+Requires CUDA with BF16 support.
 
 ```bash
-pip install torch==2.11.0 transformers==5.17.0 huggingface_hub==1.33.0
-hf download SkillScreen/SkillScreen-2B-SBR --local-dir skillscreen
+pip install huggingface_hub
+hf download SkillScreen/SkillScreen-2B-SBR --revision v1 --local-dir skillscreen
+pip install -r skillscreen/requirements.txt
 python skillscreen/score.py path/to/SKILL.md
 ```
 
-The score is the native entailment probability. A score **≥ 0.5** is flagged as suspicious. The fixed hypothesis is `This AI agent skill is malicious.` Inputs preserve the premise prefix and the complete hypothesis.
+Native maliciousness score **≥ 0.5** is flagged as suspicious.
 
-[Project page](https://jerroldh.github.io/SkillScreen-webpage/) · [Code](https://github.com/JerroldH/SkillScreen)
+[Checkpoint details](provenance.json) · [Export verification](merge_verification.json) · [Project page](https://jerroldh.github.io/SkillScreen-webpage/) · [Code](https://github.com/JerroldH/SkillScreen)
 
-This release contains one checkpoint. Paper aggregate results cover five training seeds. BF16 merging changes one of 973 validation decisions at 0.5. The original adapter is retained unchanged. Scores are uncalibrated, and a low score does not certify a package as safe. See `provenance.json` and `merge_verification.json` for the checkpoint and export checks.
+Jerrold Huang, Junwei Quan, Saiyue Lyu, Zihan Guo, Yingqiang Ge.
 
-Authors: Jerrold Huang, Junwei Quan, Saiyue Lyu, Zihan Guo, Yingqiang Ge.
-
-SkillScreen-authored material is subject to `LICENSE`. Upstream rights and notices remain applicable.
+[License](LICENSE) · [Third-party notices](THIRD_PARTY_NOTICES.md)