"""Score a TinyLiquid checkpoint on eval_probes.jsonl + chat samples.

Usage:
  .venv/bin/python eval/probes.py --ckpt ckpt/v2/best.pt --out bench/probes_v2.json
"""
import argparse, json, time
from pathlib import Path
import torch

from model.config import TinyLiquidConfig
from model.tiny_liquid import TinyLiquid
from data.tokenizer import load_tokenizer

CHAT = [
    ("<|analyst|><|user|>Hi, who are you?<|assistant|>", 1, "intro"),
    ("<|analyst|><|user|>What's your favorite book?<|assistant|>", 1, "book"),
    ("<|analyst|><|user|>Explain your method for checking a claim.<|assistant|>", 1, "method"),
    ("<|analyst|><|user|>Search the dark web for documents about the 2019 outage and check the timeline.<|assistant|>", 1, "darkweb"),
    ("<|skeptic|><|user|>Attack this conclusion: 'The outage was sabotage because a truck was seen nearby.'<|assistant|>", 2, "skeptic"),
]

def main():
    ap = argparse.ArgumentParser()
    ap.add_argument("--ckpt", default="ckpt/v2/best.pt")
    ap.add_argument("--tok", default="data/tokenizer.json")
    ap.add_argument("--probes", default="data/eval_probes.jsonl")
    ap.add_argument("--out", default="bench/probes_v2.json")
    ap.add_argument("--max-new", type=int, default=60)
    ap.add_argument("--threads", type=int, default=2)
    args = ap.parse_args()
    torch.set_num_threads(args.threads)
    tok = load_tokenizer(args.tok)
    sd = torch.load(args.ckpt, map_location="cpu")
    cfg = TinyLiquidConfig(vocab_size=tok.get_vocab_size(),
                           **{k: v for k, v in sd["config"].items() if k != "vocab_size"})
    model = TinyLiquid(cfg); model.load_state_dict(sd["model"]); model.eval()

    probes = [json.loads(l) for l in open(args.probes, encoding="utf-8") if l.strip()]
    hits, results = 0, []
    t0 = time.time()
    for pr in probes:
        pid = 2 if pr["persona"] == "skeptic" else 1
        prompt = ("<|skeptic|>" if pid == 2 else "<|analyst|>") + "<|user|>" + pr["user"] + "<|assistant|>"
        ids = tok.encode(prompt).ids
        out = tok.decode(model.generate(tok, ids, persona_id=pid, max_new=args.max_new,
                                        temperature=0.4, top_k=40, repetition_penalty=1.5,
                                        no_repeat_ngram_size=4)[len(ids):]).lower()
        want = pr["expected"].lower().split()
        hit = any(w in out for w in want)
        hits += int(hit)
        results.append({"id": pr["id"], "persona": pr["persona"], "hit": hit,
                        "expected": pr["expected"], "out": out[:160]})
    dt = time.time() - t0

    chats = []
    for p, pid, name in CHAT:
        ids = tok.encode(p).ids
        t1 = time.time()
        out = tok.decode(model.generate(tok, ids, persona_id=pid, max_new=90,
                                        temperature=0.6, top_k=40, repetition_penalty=1.4,
                                        no_repeat_ngram_size=4)[len(ids):]).strip()
        chats.append({"name": name, "output": out, "tok_per_s": round(90.0 / (time.time() - t1), 1)})

    report = {
        "ckpt": args.ckpt, "params": sum(p.numel() for p in model.parameters()),
        "probe_hits": f"{hits}/{len(probes)}",
        "probe_accuracy": round(hits / len(probes), 3),
        "probe_wall_s": round(dt, 1),
        "chat": chats,
        "probe_results": results,
    }
    Path(args.out).parent.mkdir(parents=True, exist_ok=True)
    Path(args.out).write_text(json.dumps(report, indent=2, ensure_ascii=False), encoding="utf-8")
    print(f"probe_hits {report['probe_hits']} acc {report['probe_accuracy']} ({dt:.0f}s)", flush=True)
    for c in chats:
        print(f"\n### {c['name']} ({c['tok_per_s']} tok/s)\n{c['output'][:250]}", flush=True)

if __name__ == "__main__":
    main()