File size: 5,577 Bytes
bc7101b
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
"""Quick test: does our trained model still get GSM8K answers right when forced
to emit ONLY the final answer (no verbal CoT in y)?

This is the *latent-reasoning value proposition* test: if the K=16 latent
vectors actually carry the reasoning, the model should be able to produce
"#### N" directly without writing out steps.

We don't retrain — we just generate with max_new_tokens small enough that
no verbal CoT can fit. Two flavors:
  short-32:  max_new_tokens=32, just enough for "#### NUMBER" + a few extras
  short-8:   max_new_tokens=8, basically forces direct emission

The model was trained to emit a full ~150-token GSM8K answer, so we expect
this is way out-of-distribution and accuracy will be low. The point is to
calibrate: is there ANY chance z carries the reasoning, or is the model
hopelessly dependent on writing verbal CoT?
"""
from __future__ import annotations

import argparse
import json
import re
import time
from pathlib import Path

import torch
from torch.utils.data import DataLoader

from ..data import GSM8KDataset, collate_batch
from ..model import BLTConfig, LatentProjector, build_base, generate_with_latent


GSM8K_NUM = re.compile(r"####\s*(-?\d+(?:\.\d+)?)")
ANY_NUM = re.compile(r"-?\d+(?:\.\d+)?")


def parse_pred(text: str):
    m = GSM8K_NUM.search(text)
    if m: return m.group(1)
    nums = ANY_NUM.findall(text)
    return nums[-1] if nums else None


def correct(pred, gold):
    if pred is None: return False
    try: return abs(float(pred) - float(gold)) < 1e-4
    except ValueError: return False


def main():
    p = argparse.ArgumentParser()
    p.add_argument("--ckpt", required=True)
    p.add_argument("--config", required=True)
    p.add_argument("--n", type=int, default=100)
    p.add_argument("--K", type=int, default=16)
    p.add_argument("--no_block_y_to_x", action="store_true")
    p.add_argument("--max_new_tokens", type=int, default=32)
    p.add_argument("--out", default=None)
    args = p.parse_args()

    with open(args.config) as f:
        cfg = json.load(f)

    device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
    ckpt = Path(args.ckpt)
    bcfg = BLTConfig(
        base_model=cfg["base_model"], use_lora=False,
        lora_r=cfg["lora_r"], lora_alpha=cfg["lora_alpha"],
        lora_dropout=cfg["lora_dropout"],
        lora_target_modules=tuple(cfg["lora_target_modules"]),
        K_latents=args.K, block_y_to_x=cfg["block_y_to_x"],
        proj_init_scale=cfg["proj_init_scale"],
        dtype=cfg["dtype"], attn_impl=cfg["attn_impl"],
        gradient_checkpointing=False,
    )
    base_model, tokenizer = build_base(bcfg)
    from peft import PeftModel
    adapter_dir = ckpt / "model"
    if (adapter_dir / "adapter_config.json").exists():
        model = PeftModel.from_pretrained(base_model, str(adapter_dir))
        print(f"[load] adapter from {adapter_dir}")
    else:
        model = base_model
    model.to(device).eval()
    inner = model.get_base_model() if hasattr(model, "get_base_model") else model
    d_model = inner.config.hidden_size
    projector = LatentProjector(
        d_model, init_scale=cfg["proj_init_scale"],
        use_mlp=cfg.get("proj_mlp", False),
        hidden_mult=cfg.get("proj_hidden_mult", 4),
    ).to(device).to(next(model.parameters()).dtype)
    projector.load_state_dict(torch.load(ckpt / "projector.pt", map_location=device))
    projector.eval()

    val_ds = GSM8KDataset(split="test", max_examples=args.n)
    loader = DataLoader(
        val_ds, batch_size=8, shuffle=False,
        collate_fn=lambda b: collate_batch(b, tokenizer,
                                            max_prompt_len=cfg["max_prompt_len"],
                                            max_answer_len=cfg["max_answer_len"]),
    )

    block_y_to_x = not args.no_block_y_to_x
    print(f"[mode] block_y_to_x={block_y_to_x}  max_new_tokens={args.max_new_tokens}  K={args.K}")

    correct_n = 0
    total = 0
    examples = []
    t0 = time.time()
    for batch in loader:
        x_ids = batch.x_ids.to(device); x_attn = batch.x_attn.to(device)
        B = x_ids.size(0)
        gen = generate_with_latent(
            model, tokenizer, projector,
            x_ids=x_ids, x_attn=x_attn, K=args.K,
            block_y_to_x=block_y_to_x,
            max_new_tokens=args.max_new_tokens,
            temperature=0.0, eos_token_id=tokenizer.eos_token_id,
        )
        for b in range(B):
            text = tokenizer.decode(gen[b], skip_special_tokens=True)
            pred = parse_pred(text)
            gold = batch.final_strs[b].replace("#### ", "").strip()
            ok = correct(pred, gold)
            correct_n += int(ok)
            total += 1
            if len(examples) < 8:
                examples.append({"text": text[:120], "pred": pred, "gold": gold, "ok": ok})
    summary = {
        "ckpt": str(ckpt), "n": args.n, "K": args.K,
        "block_y_to_x": block_y_to_x,
        "max_new_tokens": args.max_new_tokens,
        "acc": correct_n / max(total, 1),
        "correct": correct_n,
        "total": total,
        "elapsed_s": time.time() - t0,
        "examples": examples,
    }
    out = args.out or str(ckpt / f"short_y_eval_M{args.max_new_tokens}_block{block_y_to_x}.json")
    Path(out).write_text(json.dumps(summary, indent=2))
    print(f"[done] acc={summary['acc']:.4f} ({correct_n}/{total})  elapsed={summary['elapsed_s']:.0f}s")
    print(f"[written] {out}")
    for e in examples[:5]:
        print(f"  text={e['text']!r} pred={e['pred']} gold={e['gold']} ok={e['ok']}")


if __name__ == "__main__":
    main()