File size: 10,142 Bytes
90d6e8c
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
import json, os, time, re
import torch
from transformers import AutoModelForCausalLM, AutoModelForSequenceClassification, AutoTokenizer
from peft import PeftModel
from datasets import load_dataset

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}
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 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):
    tok = AutoTokenizer.from_pretrained('Qwen/Qwen3-4B', trust_remote_code=True)
    if tok.pad_token is None:
        tok.pad_token = tok.eos_token
    model = AutoModelForSequenceClassification.from_pretrained('Qwen/Qwen3-4B', num_labels=1, torch_dtype=torch.bfloat16, device_map='auto', trust_remote_code=True)
    model.config.pad_token_id = tok.pad_token_id
    model = PeftModel.from_pretrained(model, adapter_id)
    model.eval()
    return model, tok

def parse_action(text):
    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()
    for a in ACTIONS:
        if a in text.lower():
            return a
    return 'tool_call'

def build_proposer_messages(example):
    msgs = example['messages']
    context = '\n'.join(f"{m['role']}: {str(m['content'])[:300]}" for m in msgs[-4:])
    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'):
    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)
    return model(**inputs).logits.squeeze().item()

@torch.no_grad()
def acc_rej(model, tokenizer, proposed_action, example_msgs, device='cuda'):
    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)
    resp = tokenizer.decode(outputs[0][inputs['input_ids'].shape[1]:], skip_special_tokens=True).strip().lower()
    return 'accept' in resp and 'reject' not in resp

def build_reward_text(proposed_action, example):
    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}"

def main():
    device = 'cuda'
    print(f'Device: {device}')
    ds = load_dataset(EVAL_DS, split='train')
    data = [ds[i] for i in range(min(MAX_EVAL, len(ds)))]
    print(f'Evaluating {len(data)} examples')

    print('Loading cheap model...')
    cm, ctok = load_lm(f'{HUB_ORG}/speculative-proposer-qwen3-1.7b', device)
    print('Loading reward verifier...')
    vm, vtok = load_reward_model(f'{HUB_ORG}/speculative-verifier-qwen3-4b', device)
    print('Loading strong model...')
    sm, stok = load_lm('Qwen/Qwen3-8B', device)

    all_metrics = {}

    # A
    print('\nConfig A: strong only')
    res = []
    for i, ex in enumerate(data):
        if i % 20 == 0: print(f'  {i}/{len(data)}')
        pred = predict_action(sm, stok, build_proposer_messages(ex), device)
        res.append({'pred': pred, 'true': ex['action_type'], 'cost': COST['strong']})
    acc = sum(1 for r in res if r['pred'] == r['true']) / len(res)
    all_metrics['A'] = {'config': 'A', 'accuracy': round(acc, 4), 'avg_cost': COST['strong'], 'n': len(res)}
    print(f'  Acc: {acc:.3f}')

    # B
    print('\nConfig B: cheap only')
    res = []
    for i, ex in enumerate(data):
        if i % 20 == 0: print(f'  {i}/{len(data)}')
        pred = predict_action(cm, ctok, build_proposer_messages(ex), device)
        res.append({'pred': pred, 'true': ex['action_type'], 'cost': COST['cheap']})
    acc = sum(1 for r in res if r['pred'] == r['true']) / len(res)
    all_metrics['B'] = {'config': 'B', 'accuracy': round(acc, 4), 'avg_cost': COST['cheap'], 'n': len(res)}
    print(f'  Acc: {acc:.3f}')

    # C
    print('\nConfig C: cheap + strong verifier')
    res = []
    for i, ex in enumerate(data):
        if i % 20 == 0: print(f'  {i}/{len(data)}')
        cheap_pred = predict_action(cm, ctok, build_proposer_messages(ex), device)
        accepted = acc_rej(sm, stok, cheap_pred, ex['messages'], device)
        if accepted:
            pred, cost = cheap_pred, COST['cheap'] + COST['verify_check']
        else:
            pred = predict_action(sm, stok, build_proposer_messages(ex), device)
            cost = COST['cheap'] + COST['verify_check'] + COST['strong']
        res.append({'pred': pred, 'true': ex['action_type'], 'cost': cost, 'accepted': accepted})
    acc = sum(1 for r in res if r['pred'] == r['true']) / len(res)
    ar = sum(1 for r in res if r['accepted']) / len(res)
    all_metrics['C'] = {'config': 'C', 'accuracy': round(acc, 4), 'avg_cost': round(sum(r['cost'] for r in res)/len(res), 4), 'accept_rate': round(ar, 4), 'n': len(res)}
    print(f'  Acc: {acc:.3f} | Accept: {ar:.3f}')

    # D
    THR = -1.0
    print(f'\nConfig D: cheap + reward verifier (thr={THR})')
    res = []
    for i, ex in enumerate(data):
        if i % 20 == 0: print(f'  {i}/{len(data)}')
        cheap_pred = predict_action(cm, ctok, build_proposer_messages(ex), device)
        score = get_reward_score(vm, vtok, build_reward_text(cheap_pred, ex), device)
        res.append({'pred': cheap_pred, 'true': ex['action_type'], 'cost': COST['cheap'] + COST['verify_check'], 'accepted': score >= THR, 'score': score})
    acc = sum(1 for r in res if r['pred'] == r['true']) / len(res)
    ar = sum(1 for r in res if r['accepted']) / len(res)
    scores = [r['score'] for r in res]
    all_metrics['D'] = {'config': 'D', 'accuracy': round(acc, 4), 'avg_cost': round(sum(r['cost'] for r in res)/len(res), 4), 'accept_rate': round(ar, 4), 'mean_score': round(sum(scores)/len(scores), 3), 'min_score': round(min(scores), 3), 'max_score': round(max(scores), 3), 'n': len(res)}
    print(f'  Acc: {acc:.3f} | Accept: {ar:.3f} | Score: {sum(scores)/len(scores):.3f}')

    # E
    N = 3
    print(f'\nConfig E: multi-proposal (n={N})')
    res = []
    for i, ex in enumerate(data):
        if i % 10 == 0: print(f'  {i}/{len(data)}')
        msgs = build_proposer_messages(ex)
        text = ctok.apply_chat_template(msgs, tokenize=False, add_generation_prompt=True)
        proposals = []
        for _ in range(N):
            inputs = ctok(text, return_tensors='pt', truncation=True, max_length=2048).to(device)
            outputs = cm.generate(**inputs, max_new_tokens=50, do_sample=True, temperature=0.8, top_p=0.95, pad_token_id=ctok.pad_token_id)
            proposals.append(parse_action(ctok.decode(outputs[0][inputs['input_ids'].shape[1]:], skip_special_tokens=True)))
        scored = [(p, get_reward_score(vm, vtok, build_reward_text(p, ex), device)) for p in set(proposals)]
        best = max(scored, key=lambda x: x[1])[0]
        res.append({'pred': best, 'true': ex['action_type'], 'cost': COST['cheap'] * N + COST['verify_check'] * N})
    acc = sum(1 for r in res if r['pred'] == r['true']) / len(res)
    all_metrics['E'] = {'config': 'E', 'accuracy': round(acc, 4), 'avg_cost': round(sum(r['cost'] for r in res)/len(res), 4), 'n': len(res)}
    print(f'  Acc: {acc:.3f}')

    print(f'\n{"="*60}')
    print(f'{"Config":<6} {"Acc":>8} {"Cost":>8} {"Accept%":>10}')
    print('-'*60)
    for c in ['A','B','C','D','E']:
        m = all_metrics[c]
        ar = f'{m.get("accept_rate","-"):.3f}' if isinstance(m.get('accept_rate'), float) else '-'
        print(f'{c:<6} {m["accuracy"]:>8.3f} {m["avg_cost"]:>8.3f} {ar:>10}')

    print('\nCOST-QUALITY FRONTIER')
    for m in sorted(all_metrics.values(), key=lambda x: x['avg_cost']):
        print(f"  {m['config']}: cost={m['avg_cost']:.3f} acc={m['accuracy']:.3f}")

    with open('/tmp/eval_results_v2.json', 'w') as f:
        json.dump({'metrics': all_metrics, 'version': 'v2-fixed-prompts', 'n': len(data)}, f, indent=2)

    from huggingface_hub import HfApi
    api = HfApi()
    api.upload_file(path_or_fileobj='/tmp/eval_results_v2.json', path_in_repo='eval_results_v2.json', repo_id=f'{HUB_ORG}/speculative-tool-actions', repo_type='model', commit_message='Eval v2 results (fixed prompt format)')
    print('Uploaded!')

if __name__ == '__main__':
    main()