File size: 6,057 Bytes
aeaff79
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
#!/usr/bin/env python3
# /// script
# dependencies = ["torch", "transformers", "accelerate"]
# ///
"""Proper eval of sakthai-plus models on sakthai-bench-v2.

Uses tokenizer.apply_chat_template with tools= (identical to training rendering).
Env: MODEL, SAMPLE, BATCH, MAX_NEW, DUMP
"""
import os, json, re, time, collections, urllib.request
import torch
from transformers import AutoModelForCausalLM, AutoTokenizer

MODEL = os.environ.get("MODEL", "Nanthasit/sakthai-plus-1.5b")
BATCH = int(os.environ.get("BATCH", "1"))
SAMPLE = int(os.environ.get("SAMPLE", "50"))
MAX_NEW = int(os.environ.get("MAX_NEW", "128"))
DUMP = int(os.environ.get("DUMP", "3"))

URL = "https://huggingface.co/datasets/Nanthasit/sakthai-bench-v2/resolve/main/data/test.jsonl"
print(f"Loading {URL} ...", flush=True)
with urllib.request.urlopen(URL) as f:
    TEST = [json.loads(line) for line in f.read().decode().strip().splitlines()]
if SAMPLE:
    import random
    random.seed(42)
    TEST = random.sample(TEST, min(SAMPLE, len(TEST)))
print(f"Loaded {len(TEST)} test rows", flush=True)

def parse_tool_calls(text):
    calls = []
    for m in re.finditer(r'<tool_call>\s*(.*?)\s*</tool_call>', text, re.DOTALL):
        try:
            obj = json.loads(m.group(1))
            args = obj.get("arguments", {})
            if isinstance(args, str):
                try:
                    args = json.loads(args)
                except json.JSONDecodeError:
                    pass
            calls.append({"name": obj.get("name", ""), "arguments": args})
        except json.JSONDecodeError:
            pass
    return calls

def norm_args(a):
    if isinstance(a, str):
        try:
            a = json.loads(a)
        except json.JSONDecodeError:
            return str(a)
    if isinstance(a, dict):
        return {k: norm_args(v) for k, v in sorted(a.items()) if v is not None}
    return a

def norm_call(c):
    return {"name": c.get("name", ""), "arguments": norm_args(c.get("arguments", {}))}

def match_score(gold_calls, pred_calls):
    gs = {json.dumps(norm_call(c), sort_keys=True) for c in gold_calls}
    ps = {json.dumps(norm_call(c), sort_keys=True) for c in pred_calls}
    if not gs and not ps:
        return True, True
    correct = gs == ps
    args_ok = all(any(g["name"] == p["name"] and g["arguments"] == p["arguments"]
                      for p in pred_calls) for g in gold_calls) if pred_calls else False
    return correct, args_ok

device = "cuda" if torch.cuda.is_available() else "cpu"
print(f"Device: {device}", flush=True)

print(f"Loading tokenizer {MODEL} ...", flush=True)
tokenizer = AutoTokenizer.from_pretrained(MODEL)
if tokenizer.pad_token is None:
    tokenizer.pad_token = tokenizer.eos_token
tokenizer.padding_side = "left"

print(f"Loading model {MODEL} ...", flush=True)
model = AutoModelForCausalLM.from_pretrained(
    MODEL,
    torch_dtype=torch.float16,
    device_map="auto" if device == "cuda" else None,
    low_cpu_mem_usage=True,
).to(device)
model.eval()
print("Model loaded", flush=True)

results = collections.defaultdict(lambda: {"sel": [], "args": [], "strict": []})
held_results = collections.defaultdict(lambda: {"sel": [], "args": [], "strict": []})

t0 = time.time()
for i in range(0, len(TEST), BATCH):
    batch = TEST[i:i + BATCH]
    prompts = []
    for row in batch:
        msgs = list(row.get("messages", []))
        while msgs and msgs[-1].get("role") in ("assistant", "tool"):
            msgs.pop()
        prompts.append(tokenizer.apply_chat_template(
            msgs, tools=row.get("tools") or None,
            tokenize=False, add_generation_prompt=True,
        ))

    inputs = tokenizer(prompts, return_tensors="pt", padding=True, truncation=True, max_length=2048).to(device)
    with torch.no_grad():
        outputs = model.generate(
            **inputs,
            max_new_tokens=MAX_NEW,
            do_sample=False,
            pad_token_id=tokenizer.pad_token_id,
        )

    for j, row in enumerate(batch):
        input_len = inputs["input_ids"].shape[1]
        gen = tokenizer.decode(outputs[j][input_len:], skip_special_tokens=True)
        pred_calls = parse_tool_calls(gen)
        gold_calls = row.get("gold_calls", [])
        category = row.get("category", "unknown")
        held = row.get("held_out_tool", False)

        correct, args_ok = match_score(gold_calls, pred_calls)
        target = held_results if held else results
        target[category]["sel"].append(correct)
        target[category]["args"].append(args_ok)
        target[category]["strict"].append(correct and args_ok)

        if DUMP and j < DUMP:
            print(f"\n--- Row {i + j} ({category}) ---", flush=True)
            print(f"GOLD: {gold_calls}", flush=True)
            print(f"PRED: {pred_calls}", flush=True)
            print(f"RAW: {gen[:200]!r}", flush=True)
            print(f"CORRECT: {correct}", flush=True)

    elapsed = time.time() - t0
    print(f"  [{i + len(batch)}/{len(TEST)}] {elapsed:.0f}s elapsed", flush=True)

print("\n" + "=" * 60)
print("RESULTS")
print("=" * 60)
all_sel, all_args, all_strict = [], [], []
for cat in sorted(results):
    r = results[cat]
    n = len(r["sel"])
    sel = sum(r["sel"]) / n * 100 if n else 0
    args = sum(r["args"]) / n * 100 if n else 0
    strict = sum(r["strict"]) / n * 100 if n else 0
    all_sel.extend(r["sel"]); all_args.extend(r["args"]); all_strict.extend(r["strict"])
    print(f"  {cat:20s}  selection={sel:.1f}  arguments={args:.1f}  strict={strict:.1f}  n={n}")

if all_sel:
    n = len(all_sel)
    print(f"\n  {'AVERAGE':20s}  selection={sum(all_sel)/n*100:.1f}  arguments={sum(all_args)/n*100:.1f}  strict={sum(all_strict)/n*100:.1f}  n={n}")

hs, ha, hst = [], [], []
for cat in sorted(held_results):
    r = held_results[cat]
    hs.extend(r["sel"]); ha.extend(r["args"]); hst.extend(r["strict"])
if hs:
    n = len(hs)
    print(f"\n  {'HELD AVG':20s}  selection={sum(hs)/n*100:.1f}  arguments={sum(ha)/n*100:.1f}  strict={sum(hst)/n*100:.1f}  n={n}")

print(f"\nTotal time: {time.time() - t0:.0f}s")