File size: 14,954 Bytes
22741d9 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 | #!/usr/bin/env python3
"""Feather-specific capability scan for durable checkpoints.
This intentionally avoids transformer scale-law claims. It measures this model's own
readiness curve from checkpoints: continuation BPB, forced-choice cloze accuracy,
factual rank, exact-ish BLEU/ROUGE, and generation hygiene.
Non-invasive: reads a local checkpoint or downloads one from the Hub; never touches a
running HF Job pod.
"""
from __future__ import annotations
import argparse
import json
import math
import os
import re
import sys
import time
from collections import Counter
from pathlib import Path
from typing import Iterable
import torch
try:
sys.stdout.reconfigure(line_buffering=True) # type: ignore[attr-defined]
except Exception:
pass
ROOT = Path(__file__).resolve().parents[1]
sys.path.insert(0, str(ROOT))
def _tokenize_words(text: str) -> list[str]:
return re.findall(r"[A-Za-z0-9']+|[^\w\s]", text.lower())
def rouge_l(pred: str, ref: str) -> float:
a, b = _tokenize_words(pred), _tokenize_words(ref)
if not a or not b:
return 0.0
prev = [0] * (len(b) + 1)
for x in a:
cur = [0]
for j, y in enumerate(b, 1):
cur.append(prev[j - 1] + 1 if x == y else max(prev[j], cur[-1]))
prev = cur
lcs = prev[-1]
prec, rec = lcs / len(a), lcs / len(b)
return 0.0 if prec + rec == 0 else 2 * prec * rec / (prec + rec)
def bleu12(pred: str, ref: str) -> float:
p, r = _tokenize_words(pred), _tokenize_words(ref)
if not p or not r:
return 0.0
scores = []
for n in (1, 2):
pc = Counter(tuple(p[i:i+n]) for i in range(max(0, len(p)-n+1)))
rc = Counter(tuple(r[i:i+n]) for i in range(max(0, len(r)-n+1)))
denom = max(1, sum(pc.values()))
hit = sum(min(c, rc[g]) for g, c in pc.items())
scores.append((hit + 1e-9) / denom)
bp = 1.0 if len(p) > len(r) else math.exp(1 - len(r) / max(1, len(p)))
return bp * math.sqrt(scores[0] * scores[1])
HELDOUT_TEXTS = [
"The capital of France is Paris, a city on the Seine known for art, science, and political history.",
"Water boils at one hundred degrees Celsius at standard atmospheric pressure.",
"Photosynthesis allows plants to convert light energy, carbon dioxide, and water into sugars and oxygen.",
"William Shakespeare wrote plays including Hamlet, Macbeth, and Romeo and Juliet.",
"The theory of evolution by natural selection is associated with Charles Darwin and Alfred Russel Wallace.",
"In computer science, a hash table stores key value pairs and uses a hash function to choose a bucket.",
]
FORCED_CHOICE = [
("The capital of France is", [" Paris", " London", " Berlin", " Rome"], 0),
("Water boils at", [" 100 degrees Celsius", " 20 degrees Celsius", " minus 10 degrees Celsius", " 1000 degrees Celsius"], 0),
("Shakespeare wrote", [" Hamlet", " The Origin of Species", " The Republic", " War and Peace"], 0),
("The theory of evolution was proposed by", [" Charles Darwin", " Isaac Newton", " Albert Einstein", " Marie Curie"], 0),
("Photosynthesis produces", [" oxygen", " iron", " salt", " plastic"], 0),
("A triangle has", [" three sides", " five sides", " seven sides", " no sides"], 0),
]
GEN_PROBES = [
("The capital of France is", "Paris."),
("Water boils at", "100 degrees Celsius."),
("Once upon a time", "there was"),
("Photosynthesis is", "the process"),
("In computer science, a hash table", "stores key value pairs."),
]
def resolve_checkpoint(args: argparse.Namespace) -> Path:
if args.ckpt:
return Path(args.ckpt).expanduser().resolve()
if args.repo_id and args.job_id:
from huggingface_hub import hf_hub_download
filename = f"jobs/{args.job_id}/{args.ckpt_name}"
print(f"[scan] downloading {args.repo_id}/{filename}")
return Path(hf_hub_download(args.repo_id, filename, repo_type="model", token=os.environ.get("HF_TOKEN")))
if args.repo_id and args.repo_path:
from huggingface_hub import hf_hub_download
print(f"[scan] downloading {args.repo_id}/{args.repo_path}")
return Path(hf_hub_download(args.repo_id, args.repo_path, repo_type="model", token=os.environ.get("HF_TOKEN")))
raise SystemExit("provide --ckpt or --repo-id with --job-id/--repo-path")
def load_model(ckpt_path: Path, device: torch.device):
if os.environ.get("HYDRA_USE_NEMOTRON", "0") == "1":
import prepare_nemotron as _p_nemo
_p_nemo.ensure_tokenizer()
try:
import subsystems.sdr_retina as _sdr_retina
_sdr_retina.build_retina()
except Exception as e:
print(f"[scan] retina build/hydrate warning: {type(e).__name__}: {e}", flush=True)
from prepare import Tokenizer
from hydra.config import PostSemClawConfig
from hydra.model import PostSemClawModel
from hydra.training import config_from_dict
tokenizer = Tokenizer.from_directory()
ckpt = torch.load(str(ckpt_path), map_location="cpu", weights_only=False)
cfg_payload = ckpt.get("config") if isinstance(ckpt, dict) else None
config = config_from_dict(cfg_payload) if isinstance(cfg_payload, dict) else PostSemClawConfig(
sequence_len=int(os.environ.get("HYDRA_SEQ_LEN", "2048")),
vocab_size=tokenizer.get_vocab_size(),
)
with torch.device("meta"):
model = PostSemClawModel(config)
model.to_empty(device=device)
state = ckpt.get("model_state_dict", ckpt)
missing, unexpected = model.load_state_dict(state, strict=False)
model.eval()
if hasattr(model, "set_bos_token_id"):
model.set_bos_token_id(tokenizer.get_bos_token_id())
meta = {
"ckpt_path": str(ckpt_path),
"step": ckpt.get("step") if isinstance(ckpt, dict) else None,
"val_bpb": ckpt.get("val_bpb") if isinstance(ckpt, dict) else None,
"missing": len(missing),
"unexpected": len(unexpected),
"config": getattr(config, "__dict__", {}),
}
return model, tokenizer, meta
def ids_for(tokenizer, text: str) -> list[int]:
ids = tokenizer.encode(text)
if not ids:
bos = tokenizer.get_bos_token_id()
ids = [bos]
return ids
@torch.no_grad()
def score_text_bpb(model, tokenizer, text: str, device: torch.device) -> float:
ids = ids_for(tokenizer, text)
if len(ids) < 2:
return float("nan")
x = torch.tensor([ids[:-1]], dtype=torch.long, device=device)
y = torch.tensor([ids[1:]], dtype=torch.long, device=device)
with torch.amp.autocast(device_type="cuda", dtype=torch.bfloat16, enabled=device.type == "cuda"):
loss = model(x, y, reduction="none").reshape(-1).float().sum().item()
return loss / (math.log(2) * max(1, len(text.encode("utf-8"))))
@torch.no_grad()
def continuation_nll(model, tokenizer, prompt: str, continuation: str, device: torch.device) -> float:
pids = ids_for(tokenizer, prompt)
cids = ids_for(tokenizer, continuation)
seq = pids + cids
if len(seq) < 2:
return float("inf")
x = torch.tensor([seq[:-1]], dtype=torch.long, device=device)
y = torch.tensor([seq[1:]], dtype=torch.long, device=device)
with torch.amp.autocast(device_type="cuda", dtype=torch.bfloat16, enabled=device.type == "cuda"):
losses = model(x, y, reduction="none").reshape(-1).float()
# Continuation labels start at index len(pids)-1.
start = max(0, len(pids) - 1)
cont = losses[start:start + len(cids)]
return float(cont.mean().item()) if cont.numel() else float("inf")
@torch.no_grad()
def _sample_next(logits: torch.Tensor, mode: str, state: dict) -> int:
z = logits.float().detach().cpu()
if mode == "greedy":
return int(z.argmax().item())
if mode == "top_k":
k = min(64, z.numel())
vals, idx = torch.topk(z / 0.8, k)
return int(idx[torch.multinomial(torch.softmax(vals, dim=-1), 1).item()].item())
if mode == "top_p":
probs = torch.softmax(z / 0.8, dim=-1)
vals, idx = torch.sort(probs, descending=True)
keep = torch.cumsum(vals, dim=-1) <= 0.92
keep[0] = True
vals, idx = vals[keep], idx[keep]
vals = vals / vals.sum()
return int(idx[torch.multinomial(vals, 1).item()].item())
if mode == "mirostat":
tau = float(state.setdefault("tau", 5.0)); eta = float(state.setdefault("eta", 0.10))
mu = float(state.setdefault("mu", 2.0 * tau))
probs = torch.softmax(z, dim=-1)
vals, idx = torch.sort(probs, descending=True)
k = max(8, min(256, int(2 ** max(1.0, min(8.0, mu)))))
vals, idx = vals[:k], idx[:k]
vals = vals / vals.sum()
j = int(torch.multinomial(vals, 1).item())
p = max(float(vals[j].item()), 1e-12)
surprise = -math.log2(p)
state["mu"] = mu - eta * (surprise - tau)
return int(idx[j].item())
raise ValueError(mode)
@torch.no_grad()
def generate_sample(model, tokenizer, prompt: str, device: torch.device, max_new: int, mode: str) -> str:
ids = ids_for(tokenizer, prompt)
max_ctx = int(getattr(getattr(model, "config", None), "sequence_len", os.environ.get("HYDRA_SEQ_LEN", "2048")))
state: dict = {}
torch.manual_seed(1234 + abs(hash((prompt, mode))) % 100000)
for _ in range(max_new):
ctx = ids[-max_ctx:]
x = torch.tensor([ctx], dtype=torch.long, device=device)
with torch.amp.autocast(device_type="cuda", dtype=torch.bfloat16, enabled=device.type == "cuda"):
logits = model(x)
ids.append(_sample_next(logits[0, -1], mode, state))
return tokenizer.decode(ids)
def generation_hygiene(text: str) -> dict[str, float]:
tail = text[-512:]
chars = list(tail)
printable = sum(c.isprintable() or c in "\n\t" for c in chars) / max(1, len(chars))
alpha_space = sum(c.isalpha() or c.isspace() or c in ".,;:'\"!?-()" for c in chars) / max(1, len(chars))
toks = _tokenize_words(tail)
rep = 0.0
if len(toks) >= 8:
grams = [tuple(toks[i:i+4]) for i in range(len(toks)-3)]
rep = 1.0 - len(set(grams)) / max(1, len(grams))
return {"printable": printable, "alpha_space": alpha_space, "repeat4": rep}
def verdict(metrics: dict) -> dict[str, object]:
bpb = metrics["heldout_bpb_mean"]
fc = metrics["forced_choice_acc"]
rouge = metrics["rouge_l_mean"]
hygiene = metrics["hygiene_mean"]
return {
"english_substrate": bpb <= 1.35 and hygiene >= 0.80,
"readable_generation": hygiene >= 0.88 and metrics["repeat4_mean"] <= 0.35,
"factual_cloze_emerging": fc >= 0.50,
"bleu_rouge_emerging": rouge >= 0.20 and metrics["bleu12_mean"] >= 0.08,
"recall_ready": fc >= 0.66 and rouge >= 0.30 and bpb <= 1.15,
}
def main() -> int:
ap = argparse.ArgumentParser()
ap.add_argument("--ckpt")
ap.add_argument("--repo-id", default=os.environ.get("HF_REPO_ID", "GAInTech/feather-pretrain-checkpoints"))
ap.add_argument("--job-id")
ap.add_argument("--repo-path")
ap.add_argument("--ckpt-name", default="latest.pt")
ap.add_argument("--device", default="cuda" if torch.cuda.is_available() else "cpu")
ap.add_argument("--max-new", type=int, default=32)
ap.add_argument("--json-out")
args = ap.parse_args()
t0 = time.time()
device = torch.device(args.device if args.device != "cuda" or torch.cuda.is_available() else "cpu")
ckpt_path = resolve_checkpoint(args)
print(f"[scan] checkpoint={ckpt_path} device={device}")
model, tokenizer, meta = load_model(ckpt_path, device)
print(f"[scan] loaded step={meta['step']} missing={meta['missing']} unexpected={meta['unexpected']}")
heldout = [score_text_bpb(model, tokenizer, t, device) for t in HELDOUT_TEXTS]
forced_rows = []
for prompt, opts, gold in FORCED_CHOICE:
scores = [continuation_nll(model, tokenizer, prompt, opt, device) for opt in opts]
pred = min(range(len(scores)), key=scores.__getitem__)
forced_rows.append({"prompt": prompt, "pred": pred, "gold": gold, "ok": pred == gold, "scores": scores, "options": opts})
gen_rows = []
for mode in ("greedy", "top_k", "top_p", "mirostat"):
for prompt, ref in GEN_PROBES:
out = generate_sample(model, tokenizer, prompt, device, args.max_new, mode)
cont = out[len(prompt):] if out.startswith(prompt) else out
h = generation_hygiene(out)
gen_rows.append({"mode": mode, "prompt": prompt, "reference": ref, "output": out, "continuation": cont, "rouge_l": rouge_l(cont, ref), "bleu12": bleu12(cont, ref), **h})
mode_stats = {}
for mode in sorted({r["mode"] for r in gen_rows}):
rows = [r for r in gen_rows if r["mode"] == mode]
mode_stats[mode] = {
"rouge_l_mean": sum(r["rouge_l"] for r in rows) / len(rows),
"bleu12_mean": sum(r["bleu12"] for r in rows) / len(rows),
"hygiene_mean": sum(r["alpha_space"] for r in rows) / len(rows),
"repeat4_mean": sum(r["repeat4"] for r in rows) / len(rows),
}
best_mode = max(
mode_stats,
key=lambda m: (mode_stats[m]["rouge_l_mean"] + mode_stats[m]["bleu12_mean"] - 0.25 * mode_stats[m]["repeat4_mean"]),
)
metrics = {
"meta": {k: v for k, v in meta.items() if k != "config"},
"heldout_bpb": heldout,
"heldout_bpb_mean": float(sum(heldout) / len(heldout)),
"forced_choice": forced_rows,
"forced_choice_acc": sum(r["ok"] for r in forced_rows) / len(forced_rows),
"generations": gen_rows,
"mode_stats": mode_stats,
"best_generation_mode": best_mode,
"rouge_l_mean": mode_stats[best_mode]["rouge_l_mean"],
"bleu12_mean": mode_stats[best_mode]["bleu12_mean"],
"hygiene_mean": mode_stats[best_mode]["hygiene_mean"],
"repeat4_mean": mode_stats[best_mode]["repeat4_mean"],
"seconds": round(time.time() - t0, 3),
}
metrics["verdict"] = verdict(metrics)
print("[CAPABILITY_SCAN_JSON] " + json.dumps(metrics, sort_keys=True))
print("\n=== SUMMARY ===")
print(f"step={meta['step']} heldout_bpb={metrics['heldout_bpb_mean']:.4f} forced_choice={metrics['forced_choice_acc']:.3f} best_mode={metrics['best_generation_mode']} rougeL={metrics['rouge_l_mean']:.3f} bleu12={metrics['bleu12_mean']:.3f} hygiene={metrics['hygiene_mean']:.3f} repeat4={metrics['repeat4_mean']:.3f}")
print("mode_stats=" + json.dumps(metrics["mode_stats"], sort_keys=True))
print("verdict=" + json.dumps(metrics["verdict"], sort_keys=True))
print("\n=== GENERATIONS ===")
for r in gen_rows:
safe = r["output"].replace("\n", "\\n")
print(f"PROMPT [{r['mode']}] {r['prompt']!r} -> {safe!r}")
if args.json_out:
Path(args.json_out).write_text(json.dumps(metrics, indent=2, sort_keys=True))
return 0
if __name__ == "__main__":
raise SystemExit(main())
|