{"data":{"kind":"file","path":"README.md","version_id":"nve4j7b0rq66cfa1npy9a5pc","entry":{"name":"README.md","path":"README.md","is_directory":false,"size":1219,"modified_at":"2026-07-19T21:56:31.249000","content_hash":"a824b040d78397303d872c0c539befaab76b84806edbe783cfcf83e634fce5dc"},"entries":[],"content":"# iol-two-pass-rl\n\nTwo-turn Qwen3.5 environment matching the IOL-AI submission protocol: bounded\ncomplete-problem analysis followed by a strict final answer for one target row.\n\n- Training with `split=\"all\"`: all 160 gold rows from the v1 training, locked,\n  and former development sets.\n- Turn budgets: 640 analysis tokens, then 256 answer tokens.\n- Main reward: `chrf_weight * chrF + (1 - chrf_weight) * exact_match`.\n- Strictly unparseable outputs receive zero; there is no separate format reward\n  and no length reward or penalty.\n- Geometric score, exact match, and chrF are reported independently as metrics.\n\nThe deliberately bounded analysis turn is not treated as a failed rollout when\nit reaches 640 tokens. A capped final turn is valid only if it satisfies the\nsame strict parser used by the competition submission.\n\nExample held-out evaluation using the same sampling temperature as training and\nsubmission:\n\n```bash\nprime eval run cardan05/iol-two-pass-rl \\\n  --model Qwen/Qwen3.5-9B \\\n  --env-args '{\"split\":\"eval\",\"eval_mode\":true}' \\\n  --num-examples 24 --rollouts-per-example 1 \\\n  --max-tokens 640 --temperature 0.8 \\\n  --sampling-args '{\"enable_thinking\":false}' \\\n  --save-results --disable-tui\n```\n","encoding":"utf-8","truncated":false,"total_bytes":1219},"status":null}