validate_exp.py

"""Validate the prompt/completion explode locally, before any paid run.

Checks:
  1. explode produces the predicted row count and respects the turn cap
  2. every prompt ends with the assistant generation marker and every completion
     is non-empty and terminated (otherwise the mask trains on nothing useful)
  3. oracle: parsing each completion recovers exactly that turn's gold tool calls
  4. no bench row and no held-out tool schema survives into training
"""
import json, re, random, hashlib, pathlib, collections, sys

random.seed(20260729)
src = pathlib.Path("job-0.5b-exp.py").read_text()
ns = {"json": json, "hashlib": __import__("hashlib")}
exec(compile(src[src.index("def _text(c):"):src.index("# ── Data: explode")], "pure", "exec"), ns)
render_chatml, _assistant_body, fingerprint = ns["render_chatml"], ns["_assistant_body"], ns["fingerprint"]

_TC = re.compile(r"<tool_call>\s*(\{.*?\})\s*</tool_call>", re.DOTALL)
MAX_TURNS = 4

def names(text):
    o = []
    for m in _TC.findall(text):
        try: o.append(json.loads(m).get("name"))
        except Exception: pass
    return [n for n in o if n]

def gold_of(msg):
    o = [(tc.get("function") or {}).get("name") for tc in (msg.get("tool_calls") or [])]
    o = [n for n in o if n]
    if not o:
        o = names(msg.get("content") or "")
    return o

train = [json.loads(l) for l in open("v7/data/train.jsonl") if l.strip()]
meta = json.load(open("bench/train_exclude_fingerprints.json"))
EXCLUDE, HELD = set(meta["fingerprints"]), set(meta["held_out_tools"])
bench_fps = {json.loads(l)["fingerprint"] for l in open("bench/data/test.jsonl") if l.strip()}

def keep(ex):
    if fingerprint(ex["messages"]) in EXCLUDE: return False
    n = {(t.get("function") or {}).get("name") or t.get("name") for t in (ex.get("tools") or [])}
    return not (n & HELD)

kept = [e for e in train if keep(e)]
fail = 0

pairs, over_cap = [], 0
for ex in kept:
    msgs, tools = ex["messages"], (ex.get("tools") or None)
    idxs = [i for i, m in enumerate(msgs) if m.get("role") == "assistant"]
    if len(idxs) > MAX_TURNS:
        idxs = sorted(random.sample(idxs, MAX_TURNS)); over_cap += 1
    for i in idxs:
        c = _assistant_body(msgs[i])
        if not c.strip(): continue
        g = gold_of(msgs[i])
        row = (render_chatml(msgs[:i], tools, add_generation_prompt=True), c + "<|im_end|>", g)
        pairs.extend([row] * (3 if len(g) > 1 else 1))   # PARALLEL_OVERSAMPLE

print(f"1. explode: {len(kept)} conversations -> {len(pairs)} pairs "
      f"({over_cap} conversations hit the {MAX_TURNS}-turn cap)")

bad_prompt = sum(1 for p, _, _ in pairs if not p.endswith("<|im_start|>assistant\n"))
bad_comp = sum(1 for _, c, _ in pairs if not c.endswith("<|im_end|>") or len(c) < 12)
print(f"2. shape: {bad_prompt} prompts missing the generation marker, "
      f"{bad_comp} completions empty or unterminated")
fail += bad_prompt + bad_comp

mismatch = sum(1 for _, c, g in pairs if names(c) != g)
print(f"3. oracle: {mismatch} completions whose parsed calls differ from gold")
fail += mismatch

leak = sum(1 for e in kept if fingerprint(e["messages"]) in bench_fps)
tool_leak = sum(1 for e in kept
                if {(t.get("function") or {}).get("name") or t.get("name")
                    for t in (e.get("tools") or [])} & HELD)
print(f"4. leakage: {leak} bench rows, {tool_leak} held-out tool schemas")
fail += leak + tool_leak

cats = collections.Counter("irrelevance" if not g else ("simple" if len(g) == 1 else "parallel")
                           for _, _, g in pairs)
tot = sum(cats.values())
print(f"5. category mix of training turns: " +
      "  ".join(f"{k} {v} ({100*v/tot:.0f}%)" for k, v in cats.most_common()))

print("\nRESULT:", "PASS" if fail == 0 else f"FAIL ({fail} problems)")
sys.exit(1 if fail else 0)