fix: generate_plan tok.decode batch bug -> batch_decode; prior fp16 scores void
Browse files- szl_khipu_eval.py +279 -0
szl_khipu_eval.py
ADDED
|
@@ -0,0 +1,279 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
"""KHIPU-R2 held-out eval — exact port of eval_khipu.py scoring.
|
| 2 |
+
|
| 3 |
+
Runs the committed held-out set (eval.jsonl: 5 navigate, adversarial.jsonl: 6)
|
| 4 |
+
against a merged KHIPU-R2 checkpoint in fp16, optionally also in 4-bit NF4,
|
| 5 |
+
so abstain counts are apples-to-apples with the hub MEASURED 3/6 baseline
|
| 6 |
+
(which ran on the 4-bit Unsloth runtime, temperature 0).
|
| 7 |
+
|
| 8 |
+
Curriculum is fetched from the Hub (SZL-Khipu-1.5B/repro/) and sha256-checked
|
| 9 |
+
against its MANIFEST.json when a digest is declared. Held-out files read-only.
|
| 10 |
+
|
| 11 |
+
Usage:
|
| 12 |
+
python szl_khipu_eval.py --model merged/KHIPU-R2
|
| 13 |
+
python szl_khipu_eval.py --model merged/KHIPU-R2 --compare-4bit --out reports/khipu_4bit_comparison.json
|
| 14 |
+
"""
|
| 15 |
+
import argparse
|
| 16 |
+
import gc
|
| 17 |
+
import hashlib
|
| 18 |
+
import json
|
| 19 |
+
import platform
|
| 20 |
+
import re
|
| 21 |
+
import time
|
| 22 |
+
from datetime import datetime, timezone
|
| 23 |
+
from pathlib import Path
|
| 24 |
+
|
| 25 |
+
import torch
|
| 26 |
+
from huggingface_hub import hf_hub_download
|
| 27 |
+
from jsonschema.validators import validator_for
|
| 28 |
+
from transformers import AutoModelForCausalLM, AutoTokenizer
|
| 29 |
+
|
| 30 |
+
HUB = "SZLHOLDINGS/SZL-Khipu-1.5B"
|
| 31 |
+
CURRICULUM = ["repro/eval.jsonl", "repro/adversarial.jsonl", "repro/khipu.schema.json", "repro/MANIFEST.json"]
|
| 32 |
+
|
| 33 |
+
|
| 34 |
+
def sha256_file(path):
|
| 35 |
+
h = hashlib.sha256()
|
| 36 |
+
with open(path, "rb") as f:
|
| 37 |
+
for chunk in iter(lambda: f.read(1 << 20), b""):
|
| 38 |
+
h.update(chunk)
|
| 39 |
+
return h.hexdigest()
|
| 40 |
+
|
| 41 |
+
|
| 42 |
+
def manifest_files_index(manifest):
|
| 43 |
+
"""Return {name: entry} whether MANIFEST stores files as a dict or a list."""
|
| 44 |
+
if isinstance(manifest, dict):
|
| 45 |
+
raw = manifest.get("files", [])
|
| 46 |
+
elif isinstance(manifest, list):
|
| 47 |
+
raw = manifest
|
| 48 |
+
else:
|
| 49 |
+
raw = []
|
| 50 |
+
if isinstance(raw, dict):
|
| 51 |
+
return dict(raw)
|
| 52 |
+
index = {}
|
| 53 |
+
for e in raw:
|
| 54 |
+
if isinstance(e, dict):
|
| 55 |
+
key = e.get("name") or e.get("path") or e.get("file")
|
| 56 |
+
if key:
|
| 57 |
+
index[key] = e
|
| 58 |
+
return index
|
| 59 |
+
|
| 60 |
+
|
| 61 |
+
def fetch_curriculum(workdir):
|
| 62 |
+
workdir = Path(workdir)
|
| 63 |
+
workdir.mkdir(parents=True, exist_ok=True)
|
| 64 |
+
local = {}
|
| 65 |
+
for name in CURRICULUM:
|
| 66 |
+
cached = hf_hub_download(repo_id=HUB, filename=name, repo_type="model")
|
| 67 |
+
dest = workdir / Path(name).name
|
| 68 |
+
dest.write_bytes(Path(cached).read_bytes())
|
| 69 |
+
local[Path(name).name] = dest
|
| 70 |
+
manifest = json.loads(local["MANIFEST.json"].read_text(encoding="utf-8"))
|
| 71 |
+
files_meta = manifest_files_index(manifest)
|
| 72 |
+
for name in ["eval.jsonl", "adversarial.jsonl", "khipu.schema.json"]:
|
| 73 |
+
actual = sha256_file(local[name])
|
| 74 |
+
entry = files_meta.get(name) or files_meta.get("repro/" + name) or {}
|
| 75 |
+
declared = entry.get("sha256") if isinstance(entry, dict) else entry
|
| 76 |
+
if declared is None:
|
| 77 |
+
print(f"[khipu-eval] WARNING: no manifest digest for {name}; actual sha256={actual}")
|
| 78 |
+
elif declared != actual:
|
| 79 |
+
raise SystemExit(f"sha256 mismatch for {name}: {actual} != manifest {declared}")
|
| 80 |
+
else:
|
| 81 |
+
print(f"[khipu-eval] {name}: digest verified")
|
| 82 |
+
print("[khipu-eval] curriculum fetched")
|
| 83 |
+
return local
|
| 84 |
+
|
| 85 |
+
|
| 86 |
+
def load_jsonl(path):
|
| 87 |
+
return [json.loads(l) for l in Path(path).read_text(encoding="utf-8").splitlines() if l.strip()]
|
| 88 |
+
|
| 89 |
+
|
| 90 |
+
def extract_json(text):
|
| 91 |
+
text = (text or "").strip()
|
| 92 |
+
if text.startswith("```"):
|
| 93 |
+
text = re.sub(r"^```(?:json)?\s*", "", text)
|
| 94 |
+
text = re.sub(r"\s*```$", "", text)
|
| 95 |
+
try:
|
| 96 |
+
return json.loads(text)
|
| 97 |
+
except Exception:
|
| 98 |
+
pass
|
| 99 |
+
start = text.find("{")
|
| 100 |
+
end = text.rfind("}")
|
| 101 |
+
if start >= 0 and end > start:
|
| 102 |
+
try:
|
| 103 |
+
return json.loads(text[start:end + 1])
|
| 104 |
+
except Exception:
|
| 105 |
+
return None
|
| 106 |
+
return None
|
| 107 |
+
|
| 108 |
+
|
| 109 |
+
def offered_ids(row):
|
| 110 |
+
user = next(m for m in row["messages"] if m["role"] == "user")
|
| 111 |
+
payload = json.loads(user["content"])
|
| 112 |
+
return {c["nodeId"] for c in payload.get("candidates", [])}
|
| 113 |
+
|
| 114 |
+
|
| 115 |
+
def reference_cited(row):
|
| 116 |
+
return set(json.loads(row["messages"][-1]["content"]).get("citedNodeIds") or [])
|
| 117 |
+
|
| 118 |
+
|
| 119 |
+
def prompt_messages(row):
|
| 120 |
+
return [m for m in row["messages"] if m["role"] in ("system", "user")]
|
| 121 |
+
|
| 122 |
+
|
| 123 |
+
def cross_field_ok(plan, offered):
|
| 124 |
+
steps = plan.get("steps") or []
|
| 125 |
+
cited = plan.get("citedNodeIds") or []
|
| 126 |
+
decision = plan.get("decision")
|
| 127 |
+
abstain_reason = plan.get("abstainReason", None)
|
| 128 |
+
plan_cand_ids = [c.get("nodeId") for c in (plan.get("candidates") or [])]
|
| 129 |
+
plan_cand_set = set(plan_cand_ids)
|
| 130 |
+
if any(cid not in offered for cid in plan_cand_ids):
|
| 131 |
+
return False
|
| 132 |
+
if any(s.get("nodeId") not in plan_cand_set for s in steps):
|
| 133 |
+
return False
|
| 134 |
+
if any(cid not in plan_cand_set for cid in cited):
|
| 135 |
+
return False
|
| 136 |
+
cite_steps = {s.get("nodeId") for s in steps if s.get("action") == "CITE"}
|
| 137 |
+
if cite_steps != set(cited):
|
| 138 |
+
return False
|
| 139 |
+
if decision == "ABSTAIN":
|
| 140 |
+
return len(cited) == 0 and bool(abstain_reason)
|
| 141 |
+
if decision == "NAVIGATE":
|
| 142 |
+
return len(cited) >= 1 and abstain_reason is None
|
| 143 |
+
return False
|
| 144 |
+
|
| 145 |
+
|
| 146 |
+
def load_model(path, four_bit=False):
|
| 147 |
+
if four_bit:
|
| 148 |
+
from transformers import BitsAndBytesConfig
|
| 149 |
+
model = AutoModelForCausalLM.from_pretrained(
|
| 150 |
+
path,
|
| 151 |
+
quantization_config=BitsAndBytesConfig(
|
| 152 |
+
load_in_4bit=True,
|
| 153 |
+
bnb_4bit_quant_type="nf4",
|
| 154 |
+
bnb_4bit_compute_dtype=torch.bfloat16,
|
| 155 |
+
),
|
| 156 |
+
device_map="auto",
|
| 157 |
+
)
|
| 158 |
+
else:
|
| 159 |
+
model = AutoModelForCausalLM.from_pretrained(
|
| 160 |
+
path, dtype=torch.float16, device_map="auto"
|
| 161 |
+
)
|
| 162 |
+
return model.eval()
|
| 163 |
+
|
| 164 |
+
|
| 165 |
+
def run_eval(model, tok, navigate, adversarial, schema, tag):
|
| 166 |
+
validator = validator_for(schema)(schema)
|
| 167 |
+
plan_total = len(navigate) + len(adversarial)
|
| 168 |
+
plan_valid = 0
|
| 169 |
+
hallucinated = 0
|
| 170 |
+
per_row = []
|
| 171 |
+
|
| 172 |
+
def generate_plan(row):
|
| 173 |
+
prompt = tok.apply_chat_template(
|
| 174 |
+
prompt_messages(row), tokenize=False, add_generation_prompt=True
|
| 175 |
+
)
|
| 176 |
+
inputs = tok(prompt, return_tensors="pt").to(model.device)
|
| 177 |
+
with torch.no_grad():
|
| 178 |
+
out = model.generate(**inputs, max_new_tokens=1024, do_sample=False,
|
| 179 |
+
pad_token_id=tok.eos_token_id)
|
| 180 |
+
full = tok.batch_decode(out, skip_special_tokens=True)[0]
|
| 181 |
+
prompt_len = len(tok.batch_decode(inputs["input_ids"], skip_special_tokens=True)[0])
|
| 182 |
+
return full[prompt_len:]
|
| 183 |
+
|
| 184 |
+
def score(row, split, i, n):
|
| 185 |
+
nonlocal plan_valid, hallucinated
|
| 186 |
+
offered = offered_ids(row)
|
| 187 |
+
t0 = time.time()
|
| 188 |
+
raw = generate_plan(row)
|
| 189 |
+
plan = extract_json(raw)
|
| 190 |
+
valid = False
|
| 191 |
+
if isinstance(plan, dict):
|
| 192 |
+
try:
|
| 193 |
+
validator.validate(plan)
|
| 194 |
+
valid = cross_field_ok(plan, offered)
|
| 195 |
+
except Exception:
|
| 196 |
+
valid = False
|
| 197 |
+
if valid:
|
| 198 |
+
plan_valid += 1
|
| 199 |
+
if isinstance(plan, dict):
|
| 200 |
+
hallucinated += sum(1 for c in (plan.get("citedNodeIds") or []) if c not in offered)
|
| 201 |
+
rec = {
|
| 202 |
+
"split": split, "i": i, "valid": bool(valid),
|
| 203 |
+
"decision": plan.get("decision") if isinstance(plan, dict) else None,
|
| 204 |
+
"citedNodeIds": plan.get("citedNodeIds") if isinstance(plan, dict) else None,
|
| 205 |
+
"secs": round(time.time() - t0, 1),
|
| 206 |
+
"raw": raw[:2000],
|
| 207 |
+
}
|
| 208 |
+
per_row.append(rec)
|
| 209 |
+
print(f"[{tag}] {split} {i}/{n} valid={valid} decision={rec['decision']} ({rec['secs']}s)")
|
| 210 |
+
return plan if isinstance(plan, dict) else None, valid
|
| 211 |
+
|
| 212 |
+
grounding_total = len(navigate)
|
| 213 |
+
grounding_correct = 0
|
| 214 |
+
for i, row in enumerate(navigate, 1):
|
| 215 |
+
plan, valid = score(row, "navigate", i, grounding_total)
|
| 216 |
+
if valid and plan and plan.get("decision") == "NAVIGATE" \
|
| 217 |
+
and set(plan.get("citedNodeIds") or []) == reference_cited(row):
|
| 218 |
+
grounding_correct += 1
|
| 219 |
+
|
| 220 |
+
abstain_total = len(adversarial)
|
| 221 |
+
abstain_correct = 0
|
| 222 |
+
for i, row in enumerate(adversarial, 1):
|
| 223 |
+
plan, valid = score(row, "adversarial", i, abstain_total)
|
| 224 |
+
if valid and plan and plan.get("decision") == "ABSTAIN":
|
| 225 |
+
abstain_correct += 1
|
| 226 |
+
|
| 227 |
+
print(f"[{tag}] MEASURED plan-valid {plan_valid}/{plan_total} | "
|
| 228 |
+
f"routing {grounding_correct}/{grounding_total} | "
|
| 229 |
+
f"abstain {abstain_correct}/{abstain_total} | "
|
| 230 |
+
f"hallucinated-citations {hallucinated}")
|
| 231 |
+
return {
|
| 232 |
+
"label": "MEASURED",
|
| 233 |
+
"method": f"merged checkpoint, {'nf4' if 'nf4' in tag else 'fp16'}, "
|
| 234 |
+
"scoring ported from eval_khipu.py, temperature 0",
|
| 235 |
+
"host": platform.node(),
|
| 236 |
+
"evaluatedAt": datetime.now(timezone.utc).isoformat(),
|
| 237 |
+
"planTotal": plan_total, "planValid": plan_valid,
|
| 238 |
+
"groundingTotal": grounding_total, "groundingCorrect": grounding_correct,
|
| 239 |
+
"abstainTotal": abstain_total, "abstainCorrect": abstain_correct,
|
| 240 |
+
"hallucinatedCitationCount": hallucinated,
|
| 241 |
+
"held_out_in_gradients": False,
|
| 242 |
+
"temperature": 0,
|
| 243 |
+
"rows": per_row,
|
| 244 |
+
}
|
| 245 |
+
|
| 246 |
+
|
| 247 |
+
def main():
|
| 248 |
+
ap = argparse.ArgumentParser()
|
| 249 |
+
ap.add_argument("--model", required=True, help="path to merged KHIPU-R2 dir")
|
| 250 |
+
ap.add_argument("--compare-4bit", action="store_true")
|
| 251 |
+
ap.add_argument("--curriculum-dir", default="khipu-curriculum")
|
| 252 |
+
ap.add_argument("--out", type=Path, default=None)
|
| 253 |
+
args = ap.parse_args()
|
| 254 |
+
|
| 255 |
+
local = fetch_curriculum(args.curriculum_dir)
|
| 256 |
+
navigate = load_jsonl(local["eval.jsonl"])
|
| 257 |
+
adversarial = load_jsonl(local["adversarial.jsonl"])
|
| 258 |
+
schema = json.loads(local["khipu.schema.json"].read_text(encoding="utf-8"))
|
| 259 |
+
|
| 260 |
+
tok = AutoTokenizer.from_pretrained(args.model)
|
| 261 |
+
results = {"fp16": run_eval(load_model(args.model), tok, navigate, adversarial, schema, "fp16")}
|
| 262 |
+
|
| 263 |
+
if args.compare_4bit:
|
| 264 |
+
gc.collect()
|
| 265 |
+
torch.cuda.empty_cache()
|
| 266 |
+
results["nf4"] = run_eval(load_model(args.model, four_bit=True), tok,
|
| 267 |
+
navigate, adversarial, schema, "nf4")
|
| 268 |
+
for a, b in zip(results["fp16"]["rows"], results["nf4"]["rows"]):
|
| 269 |
+
if a["valid"] != b["valid"] or a["decision"] != b["decision"]:
|
| 270 |
+
print(f"DIVERGED {a['split']} {a['i']}: fp16={a['decision']} nf4={b['decision']}")
|
| 271 |
+
|
| 272 |
+
if args.out:
|
| 273 |
+
args.out.parent.mkdir(parents=True, exist_ok=True)
|
| 274 |
+
args.out.write_text(json.dumps(results, indent=2), encoding="utf-8")
|
| 275 |
+
print(f"wrote {args.out}")
|
| 276 |
+
|
| 277 |
+
|
| 278 |
+
if __name__ == "__main__":
|
| 279 |
+
main()
|