LauraGG's picture
Refresh code/ with latest BLT-Reasoner sources (post-campaign)
bc7101b verified
"""Pre-registered ablation evaluation for BLT-Reasoner.
Computes GSM8K accuracy under three conditions:
A. normal-z : latents from W_proj(h_{t-1}) loop
B. random-z : latents drawn from N(0, σ²) with σ matched to mean ||z||
C. zero-z : K=0 (no latents at all; y attends directly to x, but
block_y_to_x is still on so y has no information path —
expected to be ~0% if the bottleneck is working)
H1 success: acc(A) - acc(B) >= 15pp AND acc(A) - acc(C) >= 25pp
"""
from __future__ import annotations
import argparse
import json
import re
import time
from pathlib import Path
from typing import List, Optional
import torch
from torch.utils.data import DataLoader
from .data import GSM8KDataset, MATHDataset, collate_batch, extract_final_number, extract_boxed_answer
from .model import BLTConfig, LatentProjector, build_base, forward_with_latent, generate_with_latent
def parse_pred(text: str, dataset: str = "gsm8k") -> Optional[str]:
"""Extract final answer from model output. Dataset-aware:
* gsm8k: look for "#### N", fall back to last number.
* math: look for the LAST ``\\boxed{...}`` (handles latex), fall back to last number.
"""
ds = (dataset or "gsm8k").lower()
if ds == "math":
boxed = extract_boxed_answer(text)
if boxed is not None:
return boxed.strip()
# Some models emit "= ANSWER" without a box; fall back to last number.
nums = re.findall(r"-?\d+(?:\.\d+)?", text)
return nums[-1] if nums else None
# gsm8k path
m = re.findall(r"####\s*(-?\d+(?:\.\d+)?)", text)
if m:
return m[-1].strip()
nums = re.findall(r"-?\d+(?:\.\d+)?", text)
return nums[-1] if nums else None
def _normalize_math_answer(s: str) -> str:
"""Aggressively normalize MATH-style answer strings for comparison.
Strips whitespace, LaTeX wrappers, dollar signs, common formatting noise.
Not a complete LaTeX-equivalent checker — close to but weaker than the
Hendrycks et al. evaluator. For our purposes we want a fast, deterministic
string compare that catches the common-case correctness signals.
"""
if s is None:
return ""
s = s.strip().replace(" ", "")
# Strip outer $...$ math wrappers
while s.startswith("$") and s.endswith("$") and len(s) > 2:
s = s[1:-1]
# Strip \text{...} wrappers (single layer)
s = re.sub(r"\\text\{([^{}]*)\}", r"\1", s)
# Strip leading/trailing braces
while s.startswith("{") and s.endswith("}") and len(s) > 2:
s = s[1:-1]
# Strip a trailing "." (occasionally seen)
if s.endswith("."):
s = s[:-1]
return s
def correct(pred: Optional[str], gold: str, dataset: str = "gsm8k") -> bool:
if pred is None:
return False
ds = (dataset or "gsm8k").lower()
if ds == "math":
p = _normalize_math_answer(pred)
g = _normalize_math_answer(gold)
if p == g:
return True
# Also try numeric compare for the simple case
try:
return abs(float(p) - float(g)) < 1e-4
except ValueError:
return False
# gsm8k path: numeric tolerance
try:
return abs(float(pred) - float(gold)) < 1e-4
except ValueError:
return False
def estimate_z_std(model, projector, tokenizer, val_loader, device, K) -> float:
"""Run model on a few batches to estimate the per-coordinate std of z."""
from .model import forward_with_latent
model.eval()
all_z = []
with torch.no_grad():
for i, batch in enumerate(val_loader):
if i >= 4: break
x_ids = batch.x_ids.to(device)
x_attn = batch.x_attn.to(device)
y_ids = batch.y_ids.to(device)
_, z, _ = forward_with_latent(model, x_ids, x_attn, y_ids, projector, K,
block_y_to_x=True)
all_z.append(z.float().cpu())
z_cat = torch.cat(all_z, dim=0)
return float(z_cat.std().item())
def run_condition(model, projector, tokenizer, val_loader, device, K, condition: str,
z_std: float, max_new_tokens: int, temperature: float,
dataset: str = "gsm8k",
block_y_to_x: bool = True) -> dict:
"""condition in {"normal", "random", "zero"}.
`dataset` controls parsing of the gold final answer and of the prediction:
"gsm8k" → "#### N", "math" → \\boxed{...}.
"""
inner = model.get_base_model() if hasattr(model, "get_base_model") else model
d_model = inner.config.hidden_size
correct_n = 0
total = 0
examples = []
model.eval()
for batch in val_loader:
x_ids = batch.x_ids.to(device)
x_attn = batch.x_attn.to(device)
B = x_ids.size(0)
if condition == "normal":
override_z = None
K_eff = K
elif condition == "random":
override_z = torch.randn(B, K, d_model, device=device,
dtype=next(projector.parameters()).dtype) * z_std
K_eff = K
elif condition == "zero":
override_z = torch.zeros(B, 0, d_model, device=device,
dtype=next(projector.parameters()).dtype)
K_eff = 0
else:
raise ValueError(condition)
gen = generate_with_latent(
model, tokenizer, projector,
x_ids=x_ids, x_attn=x_attn, K=K_eff,
block_y_to_x=block_y_to_x, max_new_tokens=max_new_tokens,
temperature=temperature, eos_token_id=tokenizer.eos_token_id,
override_z=override_z,
)
for b in range(B):
text = tokenizer.decode(gen[b], skip_special_tokens=True)
pred = parse_pred(text, dataset=dataset)
# GSM8K final_strs are "#### N"; MATH final_strs are the boxed value already.
raw_gold = batch.final_strs[b]
gold = raw_gold.replace("#### ", "").strip() if dataset.lower() != "math" else raw_gold.strip()
ok = correct(pred, gold, dataset=dataset)
correct_n += int(ok)
total += 1
if len(examples) < 5:
examples.append({"text": text[:200], "pred": pred, "gold": gold, "ok": ok})
return {"condition": condition, "K": K_eff, "acc": correct_n / max(total, 1),
"n": total, "correct": correct_n, "examples": examples}
# ---------------------------------------------------------------------------
# Perturbation-curve evaluation (Viteri et al. ICLR 2026 analog for continuous z).
#
# At severity p ∈ [0, 1] we replace a random ⌊p·K⌋ subset of latent positions
# with Gaussian noise matched to the observed z std. At p=0 this equals
# `condition="normal"`; at p=1 this equals `condition="random"`. Intermediate
# severities give a *continuous* load-bearing measurement that we can plot
# as a curve, the way the Markovian-LM paper reports CoT-corruption sensitivity.
# A load-bearing z shows monotone, steep descent; a decorative z (your
# Abstract-CoT failure mode) shows a flat curve.
# ---------------------------------------------------------------------------
@torch.no_grad()
def _get_z_for_batch(model, projector, x_ids, x_attn, K):
"""Compute z by running the latent loop, no backprop, no y forward."""
_, z, _ = forward_with_latent(
model, x_ids, x_attn, y_ids=None, projector=projector, K=K,
block_y_to_x=True, return_z=True,
)
return z # [B, K, d]
def _perturb_z(z: torch.Tensor, severity: float, z_std: float, seed: int) -> torch.Tensor:
"""Replace ⌊severity·K⌋ randomly-chosen latent positions per example with
Gaussian noise matched to z_std. Deterministic given seed for fair compare
across severities and conditions.
"""
B, K, d = z.shape
if severity <= 0.0:
return z
n_replace = max(1, int(round(severity * K)))
g = torch.Generator(device=z.device).manual_seed(seed)
out = z.clone()
for b in range(B):
idx = torch.randperm(K, generator=g, device=z.device)[:n_replace]
noise = torch.randn(n_replace, d, device=z.device, generator=g,
dtype=z.dtype) * z_std
out[b, idx] = noise
return out
def run_perturbation_curve(model, projector, tokenizer, val_loader, device, K,
z_std: float, severities, max_new_tokens: int,
temperature: float, seed: int = 0) -> dict:
"""For each severity p, replace fraction p of latent positions with noise
and evaluate accuracy. severities is a list of floats in [0, 1].
"""
inner = model.get_base_model() if hasattr(model, "get_base_model") else model
model.eval()
curve = []
examples_at_p100 = []
for p in severities:
correct_n, total = 0, 0
for bi, batch in enumerate(val_loader):
x_ids = batch.x_ids.to(device)
x_attn = batch.x_attn.to(device)
B = x_ids.size(0)
z = _get_z_for_batch(model, projector, x_ids, x_attn, K)
z_pert = _perturb_z(z, severity=p, z_std=z_std, seed=seed + bi)
gen = generate_with_latent(
model, tokenizer, projector,
x_ids=x_ids, x_attn=x_attn, K=K,
block_y_to_x=True, max_new_tokens=max_new_tokens,
temperature=temperature, eos_token_id=tokenizer.eos_token_id,
override_z=z_pert,
)
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 p == severities[-1] and len(examples_at_p100) < 3:
examples_at_p100.append({"text": text[:200], "pred": pred,
"gold": gold, "ok": ok})
acc = correct_n / max(total, 1)
curve.append({"severity": p, "acc": acc, "correct": correct_n, "n": total})
print(f"[perturb p={p:.2f}] acc={acc:.4f} ({correct_n}/{total})")
# Compute summary statistics: monotonicity, slope.
accs = [c["acc"] for c in curve]
n_monotone = sum(1 for i in range(len(accs) - 1) if accs[i] >= accs[i + 1])
return {
"curve": curve,
"n_pairs_monotone_decreasing": n_monotone,
"n_pairs_total": len(accs) - 1,
"acc_at_0": accs[0],
"acc_at_1": accs[-1],
"drop_0_to_1": accs[0] - accs[-1],
"examples_at_max_severity": examples_at_p100,
}
def main():
parser = argparse.ArgumentParser()
parser.add_argument("--ckpt", required=True, help="path to ckpt dir containing model/, projector.pt, head.pt")
parser.add_argument("--config", required=True)
parser.add_argument("--n", type=int, default=200)
parser.add_argument("--K", type=int, default=None, help="latent count to use (defaults to config end-of-curriculum)")
parser.add_argument("--max_new_tokens", type=int, default=256)
parser.add_argument("--temperature", type=float, default=0.0)
parser.add_argument("--out", default=None)
parser.add_argument("--perturbation_curve", action="store_true",
help="Also run a perturbation-severity sweep (Viteri-style)")
parser.add_argument("--severities", default="0.0,0.25,0.5,0.75,1.0",
help="Comma-separated severities for the perturbation curve")
parser.add_argument("--no_block_y_to_x", action="store_true",
help="EVALUATE without the y→only-z bottleneck mask (lets y "
"attend to x directly during generation). Tests "
"bottleneck-as-regularizer: does z's learned structure "
"help when the inference constraint is lifted?")
args = parser.parse_args()
with open(args.config) as f:
cfg = json.load(f)
K = args.K if args.K is not None else cfg["K_curriculum"][-1][1]
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
ckpt = Path(args.ckpt)
# Build base without LoRA wrap, then attach trained adapter from disk.
bcfg_nolora = BLTConfig(
base_model=cfg["base_model"], use_lora=False,
K_latents=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"],
)
base_model, tokenizer = build_base(bcfg_nolora)
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
print(f"[load] no adapter at {adapter_dir} (using base only)")
model.to(device).eval()
inner_base = model.get_base_model() if hasattr(model, "get_base_model") else model
d_model = inner_base.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()
dataset_name = cfg.get("dataset", "gsm8k")
if dataset_name.lower() == "math":
val_ds = MATHDataset(split="test", max_examples=args.n)
else:
val_ds = GSM8KDataset(split="test", max_examples=args.n)
val_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"],
))
z_std = estimate_z_std(model, projector, tokenizer, val_loader, device, K)
print(f"[z_std estimate] {z_std:.4f} dataset={dataset_name}")
eval_block_y_to_x = not args.no_block_y_to_x
print(f"[mode] eval_block_y_to_x={eval_block_y_to_x}")
results = {}
t0 = time.time()
for cond in ["normal", "random", "zero"]:
r = run_condition(model, projector, tokenizer, val_loader, device, K,
cond, z_std, args.max_new_tokens, args.temperature,
dataset=dataset_name, block_y_to_x=eval_block_y_to_x)
results[cond] = r
print(f"[{cond}] acc={r['acc']:.4f} ({r['correct']}/{r['n']}) elapsed={time.time()-t0:.0f}s")
summary = {
"ckpt": str(ckpt), "K": K, "n": args.n, "z_std": z_std,
"eval_block_y_to_x": eval_block_y_to_x,
"dataset": dataset_name,
"results": results,
"delta_normal_minus_random": results["normal"]["acc"] - results["random"]["acc"],
"delta_normal_minus_zero": results["normal"]["acc"] - results["zero"]["acc"],
}
success_random = summary["delta_normal_minus_random"] >= 0.15
success_zero = summary["delta_normal_minus_zero"] >= 0.25
summary["H1_supported"] = bool(success_random and success_zero)
if args.perturbation_curve:
severities = [float(s) for s in args.severities.split(",")]
print(f"[perturbation_curve] severities={severities}")
curve = run_perturbation_curve(
model, projector, tokenizer, val_loader, device, K, z_std,
severities=severities, max_new_tokens=args.max_new_tokens,
temperature=args.temperature, seed=0,
)
summary["perturbation_curve"] = curve
print(f"[perturbation_curve] acc(p=0)={curve['acc_at_0']:.3f} -> "
f"acc(p=1)={curve['acc_at_1']:.3f} drop={curve['drop_0_to_1']:.3f} "
f"monotone={curve['n_pairs_monotone_decreasing']}/{curve['n_pairs_total']}")
out = args.out or str(ckpt / "ablation.json")
with open(out, "w") as f:
json.dump(summary, f, indent=2)
print(f"[written] {out}")
print(f"H1 supported? {summary['H1_supported']} "
f"(Δ_random={summary['delta_normal_minus_random']:.3f}, "
f"Δ_zero={summary['delta_normal_minus_zero']:.3f})")
if __name__ == "__main__":
main()