---
license: mit
base_model: gpt2
datasets:
- EleutherAI/bergson-wikitext-512-chunks
library_name: transformers
pipeline_tag: text-generation
---

# GPT-2 fine-tuned on bergson-wikitext-512-chunks

GPT-2 (124M) fine-tuned on [EleutherAI/bergson-wikitext-512-chunks](https://huggingface.co/datasets/EleutherAI/bergson-wikitext-512-chunks)
(WikiText-2 pre-chunked to 512-token sequences, 4,608 train chunks) using the
[bergson](https://github.com/EleutherAI/bergson) MAGIC trainer, as the trained
model for MAGIC attribution experiments.

## Training

- 4 epochs, global batch size 64 (8x data parallel), 288 steps
- AdamW, polynomial LR schedule: lr 8e-4 (start 1e-6, end 8e-5), 25% warmup, fp32
- Loss on held-out `test[:4]` chunks: 3.22 (base gpt2: 3.62)

## Files

- Standard HF model + tokenizer files
- `bergson_config.yaml` — the fully-resolved bergson run config (all fields incl. defaults) that produced this model; rerun with `python -m bergson bergson_config.yaml`
- `optimizer.pt` — AdamW second moments (`exp_avg_sq`) at the final training
  step, in bergson's `optimizer.pt` normalizer format
  (`{"state": {idx: {"exp_avg_sq": ...}}, "param_groups": [...]}` with `idx`
  indexing deduplicated `model.named_parameters()`), for gradient
  normalization in attribution runs.