---
license: apache-2.0
tags:
  - training-data-attribution
  - influence-functions
  - interpretability
---

# Retrain bank: WikiText-2 / GPT-2, random halves, seed 1008

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

Each model is GPT-2 (`gpt2`) fine-tuned on a different random **50% (2,328 documents)**
of the 4,656-document WikiText-2 training set from
[`EleutherAI/bergson-wikitext-2-4656-chunks`](https://huggingface.co/datasets/EleutherAI/bergson-wikitext-2-4656-chunks),
following the recipe of Bae et al. 2024, *Training Data Attribution via Approximate
Unrolled Differentiation* (App. B.1). `retrained/base` is trained on the full set with
the same seed.

Five banks share the same 100 subsets (`subsets.json`) and differ only in training seed:
[1004](https://huggingface.co/datasets/EleutherAI/LDS-retrain-bank-adamw-wikitext2-N4656-bs8-seed1004), [1005](https://huggingface.co/datasets/EleutherAI/LDS-retrain-bank-adamw-wikitext2-N4656-bs8-seed1005), [1006](https://huggingface.co/datasets/EleutherAI/LDS-retrain-bank-adamw-wikitext2-N4656-bs8-seed1006), [1007](https://huggingface.co/datasets/EleutherAI/LDS-retrain-bank-adamw-wikitext2-N4656-bs8-seed1007), [1008](https://huggingface.co/datasets/EleutherAI/LDS-retrain-bank-adamw-wikitext2-N4656-bs8-seed1008). Averaging query losses over the five seeds gives the ground truth used in
the bergson replication.

## What is here

| path | what it is |
|---|---|
| `retrained/base/` | the model fine-tuned on the full training set |
| `retrained/subset_*/` | 100 models, each trained on a different random half |
| `validation.csv` | **the ground truth**: per (subset, query) change in loss for the 481 validation queries, plus the EK-FAC influence sum over the removed documents |
| `subsets.json` | which document ids each subset removed |
| `config.yaml` | the exact training configuration |
| `summary*.csv` | per-query EK-FAC LDS |

## Using it

```python
from huggingface_hub import snapshot_download
import pandas as pd

path = snapshot_download("EleutherAI/LDS-retrain-bank-adamw-wikitext2-N4656-bs8-seed1008", repo_type="dataset")

# ground truth: what training on each half 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)
```

The directory also drops straight into a bergson `validate` step as `retrained_dir`.

## Measured on the five-seed ground truth

| method | LDS (mean Spearman over 481 queries, 95% CI) |
|---|---|
| EK-FAC IF | 0.468 ± 0.015 |
| SOURCE | 0.476 ± 0.015 |

Scores were computed on the fine-tuning run in
[`EleutherAI/bergson-wikitext-2-gpt2`](https://huggingface.co/EleutherAI/bergson-wikitext-2-gpt2).

## Provenance

- optimizer `adamw` (β=(0.9, 0.999), ε=1e-8), lr `3e-5` constant, weight decay `0.01`,
  batch size `8`, `3` epochs, fp32, dropout on, seed `1008`
- corpus: WikiText-2, 4,656 training documents; 481 validation queries

Produced by [bergson](https://github.com/EleutherAI/bergson)
(`examples/replicate_bae_approx_unrolling_source/wikitext_gpt2_retrain.yaml`).