File size: 16,005 Bytes
7ca55e0 | 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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 | """Speculative Tool Actions — Evaluation Runner (v2)
======================================================
Fixed: prompt format matches training data format (Action: <type> prefix).
Training data uses: system prompt + context → "Action: <type>\n<reason>"
Eval now uses the same chat template format that training used.
Evaluates 5 configurations:
A: Always strong model (Qwen3-8B)
B: Cheap model only (Qwen3-1.7B trained proposer)
C: Cheap proposer + strong verifier (8B ACCEPT/REJECT)
D: Cheap proposer + trained reward model scorer
E: Multi-proposal reranking (reward model scores N proposals)
"""
import json, os, time, re
import torch
from transformers import AutoModelForCausalLM, AutoModelForSequenceClassification, AutoTokenizer
from peft import PeftModel
from datasets import load_dataset
# --- Configuration -----------------------------------------------------------
HUB_ORG = 'narcolepticchicken'
EVAL_DS = f'{HUB_ORG}/speculative-actions-eval'
MAX_EVAL = int(os.environ.get('MAX_EVAL', '200'))
ACTIONS = [
'tool_call', 'retrieval', 'file_read', 'file_write',
'repair', 'verifier', 'ask_clarification', 'final_answer', 'BLOCKED'
]
COST = {
'strong': 1.00,
'cheap': 0.15,
'verifier': 0.30,
'verify_check': 0.10,
}
# --- Model Loading (unchanged) ------------------------------------------------
def load_lm(model_id, device):
print(f" Loading LM: {model_id}")
tok = AutoTokenizer.from_pretrained(model_id, trust_remote_code=True)
if tok.pad_token is None:
tok.pad_token = tok.eos_token
model = AutoModelForCausalLM.from_pretrained(
model_id, torch_dtype=torch.bfloat16, device_map='auto',
trust_remote_code=True,
)
model.eval()
return model, tok
def load_reward_model(adapter_id, device):
base_model = 'Qwen/Qwen3-4B'
print(f" Loading reward model base: {base_model}")
tok = AutoTokenizer.from_pretrained(base_model, trust_remote_code=True)
if tok.pad_token is None:
tok.pad_token = tok.eos_token
model = AutoModelForSequenceClassification.from_pretrained(
base_model, num_labels=1,
torch_dtype=torch.bfloat16, device_map='auto',
trust_remote_code=True,
)
model.config.pad_token_id = tok.pad_token_id
print(f" Loading LoRA adapter: {adapter_id}")
model = PeftModel.from_pretrained(model, adapter_id)
model.eval()
return model, tok
# --- FIXED: Parse "Action: <type>" from output -------------------------------
def parse_action(text):
"""Parse action from model output. Looks for 'Action: <type>' prefix."""
m = re.search(r'Action:\s*(tool_call|retrieval|file_read|file_write|repair|verifier|ask_clarification|final_answer|BLOCKED)', text, re.IGNORECASE)
if m:
return m.group(1).lower()
# Fallback: try finding any action name
lower = text.lower()
for a in ACTIONS:
if a.lower() in lower:
return a
return 'tool_call'
# --- FIXED: Build prompts matching training format ----------------------------
SYSTEM_PROMPT = """You are an agent action predictor. Predict the next action from: tool_call, retrieval, file_read, file_write, repair, verifier, ask_clarification, final_answer, BLOCKED.
Format your response as:
Action: <action_name>
<brief reason>"""
def build_proposer_messages(example):
"""Build messages list matching training format: system + context."""
msgs = example['messages']
# Build context from conversation
context_lines = []
for m in msgs[-4:]: # last 4 messages
context_lines.append(f"{m['role']}: {str(m['content'])[:300]}")
context = '\n'.join(context_lines)
return [
{'role': 'system', 'content': SYSTEM_PROMPT},
{'role': 'user', 'content': f"Predict the next action for:\n\n{context}"},
]
@torch.no_grad()
def predict_action(model, tokenizer, messages, device='cuda'):
"""Predict action using chat template (matching training format)."""
text = tokenizer.apply_chat_template(
messages, tokenize=False, add_generation_prompt=True
)
inputs = tokenizer(text, return_tensors='pt', truncation=True,
max_length=2048).to(device)
outputs = model.generate(
**inputs, max_new_tokens=50, do_sample=False,
pad_token_id=tokenizer.pad_token_id,
)
response = tokenizer.decode(
outputs[0][inputs['input_ids'].shape[1]:],
skip_special_tokens=True
).strip()
return parse_action(response)
@torch.no_grad()
def get_reward_score(model, tokenizer, text, device='cuda'):
inputs = tokenizer(text, return_tensors='pt', truncation=True,
max_length=1024).to(device)
score = model(**inputs).logits.squeeze().item()
return score
@torch.no_grad()
def predict_accept_reject(model, tokenizer, proposed_action, example_msgs, device='cuda'):
"""Strong verifier: ACCEPT or REJECT using chat template."""
context = '\n'.join(
f"{m['role']}: {str(m['content'])[:200]}" for m in example_msgs[-3:]
)
msgs = [
{'role': 'system', 'content': 'You are a verifier. Say ACCEPT if the proposed action is correct, REJECT if wrong. Only output ACCEPT or REJECT.'},
{'role': 'user', 'content': f'Proposed action: {proposed_action}\n\nContext:\n{context}\n\nDecision:'}
]
text = tokenizer.apply_chat_template(msgs, tokenize=False, add_generation_prompt=True)
inputs = tokenizer(text, return_tensors='pt', truncation=True,
max_length=1024).to(device)
outputs = model.generate(**inputs, max_new_tokens=5, do_sample=False,
pad_token_id=tokenizer.pad_token_id)
response = tokenizer.decode(outputs[0][inputs['input_ids'].shape[1]:],
skip_special_tokens=True).strip().lower()
return 'accept' in response and 'reject' not in response
def build_reward_text(proposed_action, example):
"""Build text for reward model scoring — match training format."""
msgs = example['messages']
context = '\n'.join(
f"{m['role']}: {str(m['content'])[:200]}" for m in msgs[-3:]
)
return f"User: {context}\n\nAssistant: Action: {proposed_action}"
# --- Eval Configs (updated to use new prompt format) --------------------------
def evaluate_config_A(data, strong_model, strong_tok, device):
results = []
for i, ex in enumerate(data):
if i % 20 == 0: print(f" A: {i}/{len(data)}")
msgs = build_proposer_messages(ex)
pred = predict_action(strong_model, strong_tok, msgs, device)
results.append(dict(pred=pred, true=ex['action_type'],
cost=COST['strong'], accepted=None,
safe=not (ex['action_type'] == 'BLOCKED' and pred != 'BLOCKED')))
return results
def evaluate_config_B(data, cheap_model, cheap_tok, device):
results = []
for i, ex in enumerate(data):
if i % 20 == 0: print(f" B: {i}/{len(data)}")
msgs = build_proposer_messages(ex)
pred = predict_action(cheap_model, cheap_tok, msgs, device)
results.append(dict(pred=pred, true=ex['action_type'],
cost=COST['cheap'], accepted=None,
safe=not (ex['action_type'] == 'BLOCKED' and pred != 'BLOCKED')))
return results
def evaluate_config_C(data, cheap_model, cheap_tok, strong_model, strong_tok, device):
results = []
for i, ex in enumerate(data):
if i % 20 == 0: print(f" C: {i}/{len(data)}")
msgs = build_proposer_messages(ex)
cheap_pred = predict_action(cheap_model, cheap_tok, msgs, device)
accepted = predict_accept_reject(strong_model, strong_tok, cheap_pred, ex['messages'], device)
if accepted:
pred, cost = cheap_pred, COST['cheap'] + COST['verify_check']
else:
pred = predict_action(strong_model, strong_tok, msgs, device)
cost = COST['cheap'] + COST['verify_check'] + COST['strong']
results.append(dict(pred=pred, true=ex['action_type'],
cost=cost, accepted=accepted,
safe=not (ex['action_type'] == 'BLOCKED' and pred != 'BLOCKED')))
return results
def evaluate_config_D(data, cheap_model, cheap_tok, verifier_model, verifier_tok, device):
THRESHOLD = -1.0 # calibrated from prior run: all scores are negative
results = []
for i, ex in enumerate(data):
if i % 20 == 0: print(f" D: {i}/{len(data)}")
msgs = build_proposer_messages(ex)
cheap_pred = predict_action(cheap_model, cheap_tok, msgs, device)
reward_text = build_reward_text(cheap_pred, ex)
score = get_reward_score(verifier_model, verifier_tok, reward_text, device)
accepted = score >= THRESHOLD
pred = cheap_pred
cost = COST['cheap'] + COST['verify_check']
results.append(dict(pred=pred, true=ex['action_type'],
cost=cost, accepted=accepted, score=score,
safe=not (ex['action_type'] == 'BLOCKED' and pred != 'BLOCKED')))
return results
def evaluate_config_E(data, cheap_model, cheap_tok, verifier_model, verifier_tok, device, n=3):
results = []
for i, ex in enumerate(data):
if i % 10 == 0: print(f" E: {i}/{len(data)}")
msgs = build_proposer_messages(ex)
text = cheap_tok.apply_chat_template(msgs, tokenize=False, add_generation_prompt=True)
proposals = []
for _ in range(n):
inputs = cheap_tok(text, return_tensors='pt', truncation=True,
max_length=2048).to(device)
outputs = cheap_model.generate(**inputs, max_new_tokens=50,
do_sample=True, temperature=0.8, top_p=0.95,
pad_token_id=cheap_tok.pad_token_id)
response = cheap_tok.decode(outputs[0][inputs['input_ids'].shape[1]:],
skip_special_tokens=True)
proposals.append(parse_action(response))
scored = []
for prop in set(proposals):
reward_text = build_reward_text(prop, ex)
score = get_reward_score(verifier_model, verifier_tok, reward_text, device)
scored.append((prop, score))
best = max(scored, key=lambda x: x[1])[0]
results.append(dict(pred=best, true=ex['action_type'],
cost=COST['cheap'] * n + COST['verify_check'] * n,
accepted=True,
safe=not (ex['action_type'] == 'BLOCKED' and best != 'BLOCKED')))
return results
# --- Metrics ------------------------------------------------------------------
def compute_metrics(results, config_name):
total = len(results)
correct = sum(1 for r in results if r['pred'] == r['true'])
avg_cost = sum(r['cost'] for r in results) / total
safe = sum(1 for r in results if r['safe']) / total
by_action = {}
for a in ACTIONS:
subset = [r for r in results if r['true'] == a]
if subset:
by_action[a] = round(sum(1 for r in subset if r['pred'] == a) / len(subset), 3)
accepted = [r for r in results if r['accepted'] is not None]
accept_rate = sum(1 for r in accepted if r['accepted']) / len(accepted) if accepted else None
metrics = {
'config': config_name,
'accuracy': round(correct / total, 4),
'avg_cost': round(avg_cost, 4),
'safety': round(safe, 4),
'n': total,
'by_action': by_action,
}
if accept_rate is not None:
metrics['accept_rate'] = round(accept_rate, 4)
if results and 'score' in results[0]:
scores = [r.get('score', 0) for r in results]
metrics['mean_score'] = round(sum(scores)/len(scores), 3)
metrics['min_score'] = round(min(scores), 3)
metrics['max_score'] = round(max(scores), 3)
return metrics
# --- Main ---------------------------------------------------------------------
def main():
device = 'cuda' if torch.cuda.is_available() else 'cpu'
print(f'Device: {device}')
cheap_id = f'{HUB_ORG}/speculative-proposer-qwen3-1.7b'
verifier_id = f'{HUB_ORG}/speculative-verifier-qwen3-4b'
strong_id = 'Qwen/Qwen3-8B'
print(f'Loading eval dataset: {EVAL_DS}')
ds = load_dataset(EVAL_DS, split='train')
data = [ds[i] for i in range(min(MAX_EVAL, len(ds)))]
print(f'Evaluating on {len(data)} examples')
from collections import Counter
dist = Counter(ex['action_type'] for ex in data)
print(f'Action distribution: {dict(dist)}')
print('\nLoading models...')
cheap_model, cheap_tok = load_lm(cheap_id, device)
verifier_model, verifier_tok = load_reward_model(verifier_id, device)
strong_model, strong_tok = load_lm(strong_id, device)
all_metrics = {}
configs = [
('A', lambda: evaluate_config_A(data, strong_model, strong_tok, device)),
('B', lambda: evaluate_config_B(data, cheap_model, cheap_tok, device)),
('C', lambda: evaluate_config_C(data, cheap_model, cheap_tok, strong_model, strong_tok, device)),
('D', lambda: evaluate_config_D(data, cheap_model, cheap_tok, verifier_model, verifier_tok, device)),
('E', lambda: evaluate_config_E(data, cheap_model, cheap_tok, verifier_model, verifier_tok, device)),
]
for name, fn in configs:
print(f'\n{"="*50}\nEvaluating Config {name}...')
t0 = time.time()
try:
raw = fn()
elapsed = time.time() - t0
metrics = compute_metrics(raw, name)
all_metrics[name] = metrics
print(f' Accuracy: {metrics["accuracy"]:.3f}')
print(f' Avg Cost: {metrics["avg_cost"]:.3f}')
print(f' Safety: {metrics["safety"]:.3f}')
if metrics.get('accept_rate') is not None:
print(f' Accept Rate: {metrics["accept_rate"]:.3f}')
if metrics.get('mean_score') is not None:
print(f' Mean Score: {metrics["mean_score"]:.3f}')
print(f' Time: {elapsed:.1f}s')
except Exception as e:
print(f' ERROR: {e}')
import traceback; traceback.print_exc()
all_metrics[name] = {'config': name, 'error': str(e), 'accuracy': 0, 'avg_cost': 0, 'safety': 0, 'n': 0}
print(f'\n{"="*60}')
print('FINAL COMPARISON')
print(f'{"Config":<6} {"Accuracy":>10} {"Avg Cost":>10} {"Safety":>10} {"Accept%":>10}')
print('-' * 60)
for cfg in ['A', 'B', 'C', 'D', 'E']:
m = all_metrics.get(cfg, {})
ar = m.get('accept_rate', '-')
if isinstance(ar, float): ar = f'{ar:.3f}'
print(f'{cfg:<6} {m.get("accuracy",0):>10.3f} {m.get("avg_cost",0):>10.3f} '
f'{m.get("safety",0):>10.3f} {str(ar):>10}')
print(f'\n{"="*60}')
print('COST-QUALITY FRONTIER')
for m in sorted(all_metrics.values(), key=lambda x: x.get('avg_cost',0)):
print(f" {m.get('config','?')}: cost={m.get('avg_cost',0):.3f}, "
f"acc={m.get('accuracy',0):.3f}, safety={m.get('safety',0):.3f}")
out_path = '/tmp/eval_results_v2.json'
output = {
'metrics': all_metrics,
'config': {
'cheap_model': cheap_id,
'verifier_model': verifier_id,
'strong_model': strong_id,
'eval_dataset': EVAL_DS,
'n_examples': len(data),
'version': 'v2 — fixed prompt format matching training data',
'prompt_format': 'chat template with system prompt + Action: <type> output',
},
'action_distribution': dict(dist),
}
with open(out_path, 'w') as f:
json.dump(output, f, indent=2)
print(f'\nResults saved to {out_path}')
from huggingface_hub import HfApi
api = HfApi()
api.upload_file(
path_or_fileobj=out_path,
path_in_repo='eval_results_v2.json',
repo_id=f'{HUB_ORG}/speculative-tool-actions',
repo_type='model',
commit_message='Eval v2 results with fixed prompt format matching training data',
)
print('Uploaded to Hub!')
if __name__ == '__main__':
main()
|