EleutherAI/LDS-retrain-bank-adamw-N8k-bs256

Retrain bank: plan_adam_eps1e17_8k_bs256

This repository contains 100 fully retrained language models, not just scores.

Each model is GPT-2 (gpt2) fine-tuned on the same 8,000-document corpus with a different random 1% (80 documents) held out, from the same seed and the same data order as the base model in retrained/base. Retraining is deterministic within one environment, so the models differ only by the documents removed.

That is the expensive part of any leave-k-out attribution study, and it is reusable: a new attribution method can be evaluated against this bank without retraining anything.

What is here

path	what it is
retrained/base/	the unablated fine-tuned model
retrained/subset_*/	100 models, each missing a different 1% of the corpus
validation.csv	the ground truth: per (subset, query) change in loss caused by that removal
subsets.json	which document ids each subset removed
config.yaml	the exact training configuration
filter_proponents_*/	tail-filter results: loss change when a scorer's top-ranked 1% is removed

Using it

from huggingface_hub import snapshot_download
import pandas as pd

path = snapshot_download("EleutherAI/metasmoothness-bank-plan_adam_eps1e17_8k_bs256", repo_type="dataset")

# ground truth: what removing each subset did to each query's loss
truth = pd.read_csv(f"{path}/validation.csv")

# score your own method, then correlate its predicted influence against `diff`
# LDS = mean over queries of Spearman(predicted subset sums, measured diff)

Measured on this bank

metric	value
MAGIC LDS	0.9163
EK-FAC LDS	0.3869
metasmoothness	0.9924
tail-filter delta, MAGIC	0.07060 nats
tail-filter delta, EK-FAC	0.02923 nats
tail-filter delta, random control	0.00018 nats

LDS is the mean per-query Spearman correlation between a scorer's predicted subset influence and the measured diff. The tail-filter delta is a different question on the same bank: remove the 1% a scorer ranks most influential, retrain once, and measure the query loss change against the bank's random removals as the matched control.

Provenance

• optimizer adamw, lr 0.0002, batch size 256, 2 epochs, 63 steps, seed 42
• corpus: smollm2, 8,000 documents
• retrains for one bank all run on a single GPU type: mixing types changes the retrained models by enough to shift LDS by ~0.05, which is larger than most effects being measured.

Produced by bergson.