EleutherAI/bergson-magic-scores-gpt-2

bergson MAGIC scores — GPT-2 / wikitext-2 / random validation

Per-document MAGIC attribution scores for a GPT-2 model fine-tuned on Salesforce/wikitext wikitext-2-raw-v1 train, attributing the training-loss gradient with respect to the model's output on test[3:4] (a single test example, taken as the query).

Scores were produced by the bergson attribution toolkit's MAGIC subcommand, which back-propagates through the entire training trajectory to compute, for each training doc d, ∂(L_query)/∂(w_d) — the first-order Taylor expansion of how much removing doc d from training would change the query loss.

Files

file	description
scores.pt	torch.float32 tensor, shape (36718,). Indexed by original wikitext-2 train row position (filtered <2-token rows have score 0).
summary.csv	Final Spearman/Pearson correlation between MAGIC scores and 100-subset leave-k-out training-loss diffs at validate time.
validation.csv	Per-subset (subset, diff, score_sum) records from the leave-k-out validation.

Loading

import torch
scores = torch.load("scores.pt", map_location="cpu", weights_only=True)
# scores.shape == (36718,), one entry per original wikitext-2 train row.

Validation result

Random-strategy 100-subset leave-k-out validation, batch_size=256, num_epochs=2, polynomial LR schedule:

metric	value	p
Spearman ρ	+0.9731	2.71e-64
Pearson r	+0.9724	1.00e-63
baseline_loss	3.6840

(Sorted-strategy validation against the same scores reaches ρ = +0.9926 on the same 100 subsets.)

YAML used to generate these scores

run_path: runs/gpt2_wikitext
model: gpt2
overwrite: true

data:
  dataset: Salesforce/wikitext
  subset: wikitext-2-raw-v1
  split: "train"
  chunk_length: 512

query:
  dataset: Salesforce/wikitext
  subset: wikitext-2-raw-v1
  split: "test[3:4]"
  chunk_length: 0

distributed:
  nproc_per_node: 4
  nnode: 4

batch_size: 256
num_epochs: 2
lr_schedule:
  lr_scheduler_type: polynomial
  lr: 0.0008
  lr_start: 1e-6
  lr_end: 0.00008
  warmup_steps: 0.25

subset_strategy: random
wandb_project: magic

Saved as examples/magic/gpt2_wikitext.yaml in the bergson repo.

Run with:

bergson magic examples/magic/gpt2_wikitext.yaml