#!/usr/bin/env python3
"""Run the data-scaling ladder: identical hparams, only disfluency count varies.

Every rung is evaluated on BOTH frozen validation splits:
  V_real = pilot_v2_val_adjudicated.jsonl  (real human Switchboard -- the gate)
  V_syn  = ladder/V_syn.jsonl               (held-out synthetic -- domain-gap probe)
"""
import json, os, subprocess, sys, time

PY = "/opt/vox/venv/bin/python"
LAD = "/opt/vox/sandbox/ladder"
BENCH = "/opt/vox/sandbox/benchmarks"
V_REAL = "/opt/vox/sandbox/corpus/pilot_v2_val_adjudicated.jsonl"
V_SYN = f"{LAD}/V_syn.jsonl"
RUNGS = [("L0_baseline_3k", None),
         ("L1_5k", f"{LAD}/L1_5k_train.jsonl"),
         ("L2_10k", f"{LAD}/L2_10k_train.jsonl"),
         ("L3_full", f"{LAD}/L3_full_train.jsonl")]
MODELS = ["roberta-base", "deberta-v3-base"]
COMMON = "--seed 42 --epochs 8 --batch-size 16 --lr-encoder 3e-5 --lr-head 1e-3 --max-len 128 --dtype float32"


def sh(cmd, log):
    with open(log, "a") as f:
        f.write(f"\n$ {' '.join(cmd)}\n"); f.flush()
        return subprocess.run(cmd, stdout=f, stderr=subprocess.STDOUT).returncode


def main():
    os.makedirs(BENCH, exist_ok=True)
    summary = []
    for rname, train in RUNGS:
        for m in MODELS:
            ck = f"/opt/vox/temp/ckpt/{rname}__{m}"
            log = f"/opt/vox/temp/logs/ladder_{rname}_{m}.log"
            t0 = time.time()
            if train is None:  # L0 = the pre-Phase-C checkpoint, already trained
                src = f"/opt/vox/temp/ckpt/{m}"
                if not os.path.isdir(src):
                    print(f"skip L0 {m}: no baseline ckpt"); continue
                ck = src
            elif not os.path.exists(f"{ck}/train_meta.json"):
                rc = sh([PY, "-u", "/opt/vox/sandbox/scripts/train_span.py", "--model",
                         f"FacebookAI/roberta-base" if m == "roberta-base" else "microsoft/deberta-v3-base",
                         "--out", ck, "--train", train] + COMMON.split(), log)
                if rc != 0:
                    print(f"TRAIN FAIL {rname} {m} rc={rc}"); continue
            meta = json.load(open(f"{ck}/train_meta.json"))
            print(f"{rname:16} {m:18} n={meta['n_train']:6} loss {meta['train_log'][0]['train_loss']}"
                  f"->{meta['train_log'][-1]['train_loss']} ({meta['wall_s']}s)", flush=True)

            row = {"rung": rname, "model": m, "n_train": meta["n_train"],
                   "final_train_loss": meta["train_log"][-1]["train_loss"],
                   "train_s": meta["wall_s"]}
            for vname, vpath in (("real", V_REAL), ("syn", V_SYN)):
                ej = f"{BENCH}/ladder_{rname}_{m}_{vname}.json"
                if not os.path.exists(ej):
                    rc = sh([PY, "-u", "/opt/vox/sandbox/scripts/eval_span.py", "--model", ck,
                             "--tag", f"{rname}_{m}_{vname}", "--out", ej, "--val", vpath], log)
                    if rc != 0 or not os.path.exists(ej):
                        print(f"  EVAL FAIL {rname} {m} {vname}"); continue
                b = json.load(open(ej))["best"]
                row[f"{vname}_EM"] = b["exact_match"]
                row[f"{vname}_F1"] = b["span_word"]["f1"]
                row[f"{vname}_P"] = b["span_word"]["precision"]
                row[f"{vname}_R"] = b["span_word"]["recall"]
                row[f"{vname}_overdel"] = b["over_deletion_rate"]
                row[f"{vname}_harmful"] = b["harmful_span_rate"]
                row[f"{vname}_thr"] = b["threshold"]
                row[f"{vname}_gate"] = b["passes_gate"]
                row[f"{vname}_DIS_EM"] = b["per_category"].get("DISFLUENCY", {}).get("exact_match")
            summary.append(row)
            print(f"   REAL EM={row.get('real_EM')} F1={row.get('real_F1')} "
                  f"overdel={row.get('real_overdel')}% harmful={row.get('real_harmful')}% "
                  f"DIS_EM={row.get('real_DIS_EM')} gate={row.get('real_gate')}", flush=True)
            print(f"   SYN  EM={row.get('syn_EM')} F1={row.get('syn_F1')} "
                  f"overdel={row.get('syn_overdel')}% harmful={row.get('syn_harmful')}% "
                  f"DIS_EM={row.get('syn_DIS_EM')}", flush=True)

    json.dump(summary, open(f"{BENCH}/ladder_summary.json", "w"), indent=2)
    print("\n=== LADDER (V_real, production-representative) ===")
    print(f"{'rung':16} {'model':18} {'n':>6} {'EM':>7} {'F1':>6} {'DIS_EM':>7} "
          f"{'overdel':>8} {'harmful':>8} gate")
    for s in summary:
        print(f"{s['rung']:16} {s['model']:18} {s['n_train']:>6} {str(s.get('real_EM')):>7} "
              f"{str(s.get('real_F1'))[:6]:>6} {str(s.get('real_DIS_EM')):>7} "
              f"{str(s.get('real_overdel')):>8} {str(s.get('real_harmful')):>8} {s.get('real_gate')}")


if __name__ == "__main__":
    main()