scripts/parse_rl_run.py

#!/usr/bin/env python3
"""parse_rl_run.py — extract the REAL hosted-RL evidence from a prime train run's logs.

Produces two JSON artifacts the submission folds in:
  * rl_train_curve.json — per-step dense train reward on the train pool (HumanEval 0-49).
    This is the PRIMARY "real RL, reward moved during training" evidence.
  * rl_after.json — the held-out (HumanEval 50-74) eval trajectory: base (step 0, untrained)
    vs each eval checkpoint. eval_base_model=true gives the BEFORE inside the same run, so
    base->final is a clean before->after on a disjoint split, same harness, no checkpoint to serve.

Usage:  parse_rl_run.py <run_id>
Reads `prime train logs <run_id>`; honest about whatever the numbers actually are.
"""
import json
import re
import subprocess
import sys
from pathlib import Path

RUN = sys.argv[1] if len(sys.argv) > 1 else "n8xzr01yconncax7v0i7gn0u"
RESULTS = Path(__file__).resolve().parents[1] / "results"

logs = subprocess.run(["prime", "train", "logs", RUN], capture_output=True, text=True).stdout

# --- train curve: "Step N | Time: Xs | Reward: R | Seq. Length: L" ---
steps = []
for m in re.finditer(r"Step (\d+) \| Time: ([\d.]+)s \| Reward: ([\d.]+) \| Seq\. Length: ([\d.]+)", logs):
    steps.append({"step": int(m.group(1)), "train_reward": round(float(m.group(3)), 4),
                  "time_s": round(float(m.group(2)), 2), "seq_len": round(float(m.group(4)), 1)})
steps.sort(key=lambda d: d["step"])

# --- held-out eval: "Running evals at step=N" then "Evaluated ... Avg@1=X" ---
evals = []
pending_step = None
for line in logs.splitlines():
    ms = re.search(r"Running evals at step=(\d+)", line)
    if ms:
        pending_step = int(ms.group(1))
    me = re.search(r"Evaluated art87able/spec-rl in ([\d.]+)s \(Avg@1=([\d.]+)", line)
    if me and pending_step is not None:
        evals.append({"step": pending_step, "heldout_avg_at_1": round(float(me.group(2)), 4),
                      "eval_time_s": round(float(me.group(1)), 2)})
        pending_step = None
evals.sort(key=lambda d: d["step"])

train_curve = {
    "note": "Per-step dense unit-test reward on the TRAIN pool (HumanEval 0-49) during real hosted "
            "GRPO post-training of Laguna XS.2 on art87able/spec-rl. temperature=1.0 (exploration), so "
            "step-to-step values carry sampling noise; the trajectory (not any single step) is the signal. "
            "PRIMARY evidence that we post-trained the model, not just evaluated the tool.",
    "run_id": RUN, "env": "art87able/spec-rl@0.1.5", "model": "poolside/Laguna-XS.2",
    "max_steps": 20, "batch_size": 64, "rollouts_per_example": 8, "learning_rate": 1e-6,
    "train_pool": "HumanEval/0-49", "free_hosted_train": True,
    "steps": steps,
}
after = {
    "note": "Held-out (HumanEval 50-74, DISJOINT from the train pool) eval trajectory computed BY THE "
            "TRAINER via eval_base_model=true + interval=5. step 0 = BEFORE (untrained base Laguna); later "
            "steps = AFTER checkpoints. Same harness, same hosted inference, no checkpoint served. Eval "
            "sampling carries MoE/temperature noise (see determinism_check.json), so read the trajectory.",
    "run_id": RUN, "env": "art87able/spec-rl@0.1.5", "split": "HumanEval/50-74 (held-out)",
    "model": "poolside/Laguna-XS.2",
    "before_step0_heldout": evals[0]["heldout_avg_at_1"] if evals else None,
    "after_final_heldout": evals[-1]["heldout_avg_at_1"] if evals else None,
    "delta": round(evals[-1]["heldout_avg_at_1"] - evals[0]["heldout_avg_at_1"], 4) if len(evals) >= 2 else None,
    "trajectory": evals,
    "local_greedy_before_ref": "results/rl_before.json (0.9378, greedy T=0 — corroborates step-0 base)",
}
(RESULTS / "rl_train_curve.json").write_text(json.dumps(train_curve, indent=2))
(RESULTS / "rl_after.json").write_text(json.dumps(after, indent=2))
print("train steps parsed:", len(steps))
print("  rewards:", [s["train_reward"] for s in steps])
print("held-out evals:", evals)
print("BEFORE step0:", after["before_step0_heldout"], "AFTER final:", after["after_final_heldout"], "delta:", after["delta"])