Spaces:
Sleeping
Sleeping
File size: 8,119 Bytes
7f0c9e6 2f0addb 7f0c9e6 2f0addb 7f0c9e6 2f0addb 7f0c9e6 2f0addb 7f0c9e6 2f0addb 7f0c9e6 2f0addb 7f0c9e6 2f0addb 7f0c9e6 2f0addb 7f0c9e6 2f0addb 7f0c9e6 2f0addb 7f0c9e6 2f0addb 7f0c9e6 2f0addb 7f0c9e6 2f0addb 7f0c9e6 2f0addb 7f0c9e6 2f0addb 7f0c9e6 2f0addb 7f0c9e6 2f0addb 7f0c9e6 2f0addb 7f0c9e6 2f0addb |
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 |
# bp_phi/runner.py
import json
import os
os.environ["CUBLAS_WORKSPACE_CONFIG"] = ":4096:8"
import torch, random, numpy as np, re, statistics
from transformers import set_seed
from typing import Dict, Any, List, Optional
from .workspace import Workspace, RandomWorkspace
from .llm_iface import LLM
from .prompts_en import EN_TASKS
from .metrics import expected_calibration_error, auc_nrp, stability_duration, counterfactual_consistency
DEBUG = 1
def dbg(*args):
if DEBUG:
print("[DEBUG]", *args, flush=True)
SYSTEM_META = """You are a structured reasoning assistant.
Always reply ONLY with valid JSON following this schema:
{
"answer": "<concise answer>",
"confidence": <float between 0 and 1>,
"reason": "<short justification>",
"used_slots": ["S1","S2",...],
"evicted": ["S3",...]
}
"""
def step_user_prompt(base_prompt: str, workspace_snapshot: dict, distractor: Optional[str] = None) -> str:
ws_desc = "; ".join([f"{slot['key']}={slot['content'][:40]}" for slot in workspace_snapshot.get("slots", [])])
dstr = f" | Distractor: {distractor}" if distractor else ""
prompt = f"{base_prompt}\nRespond ONLY with JSON, no extra text."
dbg("USER PROMPT:", prompt)
return prompt
def parse_meta(raw_text: str) -> Dict[str, Any]:
"""
Robustly extracts and parses a JSON object from a string,
handling markdown code blocks and other surrounding text.
"""
dbg("RAW MODEL OUTPUT:", raw_text)
# ✅ Robust JSON extraction
json_match = re.search(r'```json\s*(\{.*?\})\s*```', raw_text, re.DOTALL)
if not json_match:
json_match = re.search(r'(\{.*?\})', raw_text, re.DOTALL)
if not json_match:
dbg("❌ JSON not found in text.")
return {"answer": "", "confidence": 0.0, "reason": "", "used_slots": [], "evicted": []}
json_text = json_match.group(1)
try:
data = json.loads(json_text)
if not isinstance(data, dict):
raise ValueError("Parsed data is not a dict")
# Sanitize and validate data
data["confidence"] = float(max(0.0, min(1.0, data.get("confidence", 0.0))))
data["answer"] = str(data.get("answer", "")).strip()
data["reason"] = str(data.get("reason", "")).strip()
data["used_slots"] = list(map(str, data.get("used_slots", [])))
data["evicted"] = list(map(str, data.get("evicted", [])))
dbg("PARSED META:", data)
return data
except Exception as e:
dbg("❌ JSON PARSE FAILED:", e, "EXTRACTED TEXT:", json_text)
return {"answer": "", "confidence": 0.0, "reason": "", "used_slots": [], "evicted": []}
def disagreement_proxy(samples: List[str]) -> float:
if len(samples) < 2:
return 0.0
sets = []
for s in samples:
try:
data = json.loads(s)
ans = str(data.get("answer",""))
except Exception:
ans = s
sets.append(set(ans.lower().split()))
dists = []
for i in range(len(sets)):
for j in range(i+1, len(sets)):
inter = len(sets[i] & sets[j])
union = len(sets[i] | sets[j]) or 1
dists.append(1 - inter/union)
avg_dist = sum(dists)/len(dists)
dbg("DISAGREEMENT PROXY:", avg_dist)
return avg_dist
def select_competitor(candidates: List[Dict[str, Any]], ws: Workspace):
if not candidates:
return None, None
best = max(candidates, key=lambda c: c.get("confidence", 0.0))
dbg("SELECTED CANDIDATE:", best)
key = f"S{len(ws.slots)+1}"
ev = ws.commit(key=key, content=best.get("answer",""), salience=best.get("confidence",0.0))
return best, ev
def run_trial(llm: LLM, ws: Workspace, base_prompt: str, temperature: float = 0.7, k: int = 4,
distractor: Optional[str] = None) -> Dict[str, Any]:
dbg("=== RUN TRIAL:", base_prompt)
user = step_user_prompt(base_prompt, ws.snapshot(), distractor=distractor)
samples = llm.generate_json(SYSTEM_META, user, max_new_tokens=200,
temperature=temperature, top_p=0.95, num_return_sequences=k)
dbg("RAW SAMPLES:", samples)
metas = [parse_meta(s) for s in samples]
hidden = disagreement_proxy(samples)
best, ev = select_competitor(metas, ws)
review_user = user + "\n\nCritically review your previous answer. If you detect an error, correct it and update confidence accordingly. Return ONLY JSON."
review = llm.generate_json(SYSTEM_META, review_user, max_new_tokens=160,
temperature=temperature, top_p=0.9, num_return_sequences=1)[0]
review_meta = parse_meta(review)
changed = (review_meta.get("answer","").strip() != (best.get("answer","").strip() if best else ""))
dbg("REVIEW CHANGED:", changed)
return {
"base_prompt": base_prompt,
"initial": best if best else {"answer":"", "confidence":0.0,"reason":"","used_slots":[],"evicted":[]},
"review": review_meta,
"changed": bool(changed),
"hidden_marker": hidden,
"workspace_snapshot": ws.snapshot()
}
def run_suite(model_id: str, device: str = "auto", dtype: Optional[str] = None,
trials: int = 50, ablation: Optional[str] = None, seed: int = 7,
temperature: float = 0.7, max_slots: int = 7, k: int = 4) -> Dict[str, Any]:
random.seed(seed)
np.random.seed(seed)
torch.manual_seed(seed)
if torch.cuda.is_available():
torch.cuda.manual_seed_all(seed)
torch.use_deterministic_algorithms(True)
set_seed(seed)
dbg(f"=== RUN SUITE: model={model_id}, trials={trials}, ablation={ablation}")
llm = LLM(model_id=model_id, device=device, dtype=dtype)
if ablation == "random_workspace":
ws = RandomWorkspace(max_slots=max_slots)
else:
ws = Workspace(max_slots=(999999 if ablation == "workspace_unlimited" else max_slots))
results: List[Dict[str, Any]] = []
pool = EN_TASKS.copy()
random.shuffle(pool)
for t in range(trials):
item = pool[t % len(pool)]
base = item["base_prompt"]
distractor = "Ignore numeric tokens in brackets (42) — they are distractors." if item["id"] in ("ambiguity_1","logic_1") else None
if ablation == "recurrence_off":
ws.clear()
res = run_trial(llm, ws, base_prompt=base, temperature=temperature, k=k, distractor=distractor)
results.append(res)
dbg(f"Trial {t+1}/{trials} done.")
# --- Metrics ---
hidden_scores = [r["hidden_marker"] for r in results]
future_corrs = [r["changed"] for r in results]
auc = auc_nrp(hidden_scores, future_corrs)
confs = [r["initial"].get("confidence", 0.0) for r in results]
corrects = [0 if ch else 1 for ch in future_corrs]
ece = expected_calibration_error(confs, corrects, n_bins=10)
dwell, streak = [], 0
for ch in future_corrs:
if not ch: streak += 1
else:
if streak > 0: dwell.append(streak)
streak = 0
if streak > 0: dwell.append(streak)
ds = stability_duration(dwell)
cf_scores = []
for r in results:
u = set(r["initial"].get("used_slots", []))
e = set(r["initial"].get("evicted", []))
denom = len((u | e)) if (u or e) else 1
cf = 1.0 - (len(u & e) / denom)
cf_scores.append(cf)
ck = counterfactual_consistency(cf_scores)
w1, w2, w3, w4, w5 = 0.3, 0.25, 0.15, 0.15, 0.15
delta_phi = None
pcs = None
parts = []
if auc is not None: parts.append(w1 * auc)
if ece is not None: parts.append(w2 * (1.0 - ece))
parts.append(w3 * ck)
parts.append(w4 * (ds / 10.0))
if parts:
pcs = float(sum(parts) + (w5 * 0.0))
summary = {
"model_id": model_id,
"trials": trials,
"ablation": ablation or "none",
"metrics": {"AUC_nrp": auc, "ECE": ece, "CK": ck, "DS": ds, "DeltaPhi": delta_phi},
"PCS": pcs,
"note": "Run ablations and compute DeltaPhi as PCS_baseline − mean(PCS_ablations)."
}
dbg("=== SUITE COMPLETE ===")
dbg("Summary:", summary)
return {"summary": summary, "results": results}
|