{"data":{"kind":"file","path":"README.md","version_id":"bx6q754e5hsn402kut73kl35","entry":{"name":"README.md","path":"README.md","is_directory":false,"size":2711,"modified_at":"2026-03-14T02:41:20.953000","content_hash":"6afbe4903f49006cdc19a906e8a9523e74f5163700c654a8681f167ee9efbd43"},"entries":[],"content":"# brfauna\n\n### Overview\n- **Environment ID**: `brfauna`\n- **Short description**: Verifiers port of BRFauna eval suite. Abs: https://journals-sol.sbc.org.br/index.php/jbcs/article/view/5814\n- **Tags**: portuguese, qa, text-simplification, text-summarization, brfauna, train, eval\n\n### Datasets\n- **Primary dataset(s)**: `pira`, `porsimples`, and `recognasumm`.\n- **Source links**: \n  - [Pira](https://huggingface.co/datasets/paulopirozelli/pira)\n  - [PorSimplesSent](https://github.com/sidleal/porsimplessent)\n  - [RecognaSumm](https://huggingface.co/datasets/recogna-nlp/recognasumm)\n- **Split sizes**: \n  - `pira`: train (1806), eval/test (227)\n  - `porsimples`: train (1100), test (349)\n  - `recognasumm`: train (81k), eval (27k)\n\n### Task\n- **Type**: `vf.EnvGroup` (combines multiple `vf.SingleTurnEnv` sub-environments)\n- **Parser**: `vf.XMLParser`\n- **Rubric overview**: Combines various exact match and LLM metrics depending on the task (`judge_score_func`, `bert_score_f1_reward`, `format_reward`, `sari_reward`, `rouge_L_reward`, `meteor_reward`, `overlap_f1_reward`).\n\n### Quickstart\nRun an evaluation with default settings:\n\n```bash\nprime eval run brfauna\n```\n\nConfigure model and sampling:\n\n```bash\nprime eval run brfauna \\\n  -m gpt-4.1-mini \\\n  -n 20 -r 3 -t 1024 -T 0.7 \\\n  -a '{\"seed\": 42}'  # global args passed to all tasks\n```\n\nNotes:\n- Use `-a` / `--env-args` to pass environment-specific configuration as a JSON object.\n\n**Using Task-specific Keyword Arguments:**\nTo configure arguments specifically for one or more sub-tasks instead of globally, use the `task_kwargs` argument. It takes a dictionary mapping task names to their specific keyword arguments. These `task_kwargs` override the global kwargs for that specific task:\n\n```bash\nprime eval run brfauna \\\n  -a '{\"task_kwargs\": {\"pira\": {\"seed\": 137}, \"recognasumm\": {\"num_examples\": 100}}}'\n```\n\n### Environment Arguments\nThis environment suite accepts global arguments that are propagated to all tasks (like `seed` or `judge_model`).\n\n| Arg | Type | Default | Description |\n| --- | ---- | ------- | ----------- |\n| `tasks` | list[str] \\| str | `\"all\"` | Tasks to run (\"all\", or specific tasks like \"pira\", \"porsimples\" and \"recognasumm\") |\n| `task_kwargs` | dict \\| None | `None` | Dictionary mapping task names to task-specific arguments. Overrides global arguments. |\n\n\n### Metrics\n| Metric | Meaning |\n| ------ | ------- |\n| `reward` | The main scalar reward for each task (weighted sum) |\n| `judge_score` | LLM-as-a-judge score |\n| `bert_score_f1_reward`, `overlap_f1_reward`, `sari_reward`, `rouge_L_reward`, `meteor_reward` | Standard generative metrics depending on the specific task. |\n| `format_reward` | Formatting reward |\n","encoding":"utf-8","truncated":false,"total_bytes":2711},"status":null}