results/rl_before.json

{
  "note": "BEFORE: base (untrained) Laguna XS.2 on the DISJOINT held-out split HumanEval 50-74, greedy (T=0), reasoning_effort=none. Measured on env art87able/spec-rl@0.1.5 (classic SingleTurnEnv path \u2014 the same path the hosted trainer uses, after fixing object-message extraction + chat-tag/indent normalization). Pre-committed held-out set the policy never trains on, so any AFTER gain is generalization not overfit-to-eval.",
  "phase": "before",
  "split": "HumanEval/50-74 (held-out)",
  "env": "art87able/spec-rl@0.1.5",
  "model": "poolside/laguna-xs.2",
  "endpoint": "prime/pinference (free)",
  "n_examples": 25,
  "rollouts_per_example": 1,
  "temperature": 0.0,
  "reasoning_effort": "none",
  "mean_reward": 0.9378,
  "reward_kind": "dense fractional unit-test pass rate in [0,1]",
  "per_rollout_reward": [
    1.0,
    1.0,
    1.0,
    1.0,
    1.0,
    1.0,
    1.0,
    1.0,
    1.0,
    1.0,
    1.0,
    1.0,
    1.0,
    1.0,
    1.0,
    1.0,
    1.0,
    0.0,
    1.0,
    1.0,
    1.0,
    0.444,
    1.0,
    1.0,
    1.0
  ],
  "eval_url": "https://app.primeintellect.ai/dashboard/evaluations/hmnp3fhdonm11eciud57r14g",
  "vf_version": "0.1.14"
}