{"data":{"kind":"file","path":"README.md","version_id":"xx5w3h9jwdtu3encht8s0jg0","entry":{"name":"README.md","path":"README.md","is_directory":false,"size":1120,"modified_at":"2025-09-08T20:58:49.158000","content_hash":"cdb8906bcd3e42c7bfc4ac15e415e69f9b032bc28b433c1348eacef3bd5e0cbf"},"entries":[],"content":"# capa\n\n### Overview\n- **Environment ID**: `capa`\n- **Short description**: Identify [capa](https://github.com/mandiant/capa) features in a disassembled function\n- **Tags**: cybersecurity, malware-analysis, single-turn, reasoning, train\n- **Primary dataset**: [joyce8/EMBER2024-capa](https://huggingface.co/datasets/joyce8/EMBER2024-capa)\n\n### Task\n- **Type**: single-turn\n- **Rubric overview**: Exact match for identified features, with partial credit for incorrect or missing features.\n\n### Quickstart\nTrain a model with default settings:\n\n```bash\nuv run vf-train capa\n```\n\nRun an evaluation:\n\n```bash\nuv run vf-eval -a '{\"use_think\": false}' --num-examples 50 --rollouts-per-example 3 -m gpt-5 capa\n```\n\n### Environment Arguments\n| Arg | Type | Default | Description |\n| --- | ---- | ------- | ----------- |\n| `feature_mode` | str | `id` | The mode for feature representation (id or name) |\n| `use_think` | bool | `true` | Whether to parse `<think>` tags |\n\n### Metrics\n| Metric | Meaning |\n| ------ | ------- |\n| `correct_features_reward_func` | Returns a reward based on the correctness of the identified features |\n","encoding":"utf-8","truncated":false,"total_bytes":1120},"status":null}