anpaurehf/openwebexpertselections

OpenWebText Expert Selections (GPT-OSS 20B)

This dataset contains top-k expert selections for each token from the first 200,000,000 tokens of vietgpt/openwebtext_en, using the router logits from openai/gpt-oss-20b. Sequences are chunked to a maximum length of 32 tokens within each document (no cross-document continuity); shorter tail chunks are included without padding.

Files

• openwebtext_200m_idx.npy: uint16 indices, shape (200_000_000, 24, 4)
• openwebtext_200m_val.npy: float16 router logits, shape (200_000_000, 24, 4)
• openwebtext_200m.meta.json: run metadata (model, dataset, seq_len, topk, etc.)
• openwebtext_200m.state.json: resume state (tokens_written, dataset offsets)
• collect_experts.py: collection script used to generate the files

Notes

• Top-k values are raw router logits (pre-softmax), not probabilities.
• Tokenization uses the model tokenizer (openai/gpt-oss-20b) with add_special_tokens=False.

Evaluation (text.txt, 2,584 tokens)

• MLP (L=5, 50k steps): top1=0.4822, top5=0.6649, top10=0.7078
• Transformer v2 (L=10, 25k steps): top1=0.7663, top5=0.8266, top10=0.8402

Eval command (no generation)

uv run python generate_and_eval.py \
  --text-file text.txt \
  --max-new-tokens 0 \
  --seq-len 32 \
  --layers 10 \
  --mlp-layers 5 \
  --tx-layers 10 \
  --mlp-ckpt inverter_mlp_l5_50k.pt \
  --tx-ckpt inverter_v2_l10_accum.pt \
  --tx-d-model 768 \
  --tx-n-layer 6 \
  --tx-n-head 12 \
  --tx-d-ff 2048 \
  --layer-gating \
  --attn-impl flash_attention_2 \
  --out user_text_eval.json

Loading (memmap)

import numpy as np

idx = np.memmap(
    "openwebtext_200m_idx.npy",
    mode="r",
    dtype=np.uint16,
    shape=(200_000_000, 24, 4),
)
val = np.memmap(
    "openwebtext_200m_val.npy",
    mode="r",
    dtype=np.float16,
    shape=(200_000_000, 24, 4),
)

Collection command (for reference)

uv run python collect_experts.py \
  --dataset vietgpt/openwebtext_en \
  --model openai/gpt-oss-20b \
  --topk 4 \
  --seq-len 32 \
  --batch-size 32 \
  --max-tokens 200000000 \
  --save-every 1000000 \
  --attn-impl flash_attention_2 \
  --output openwebtext_200m.npy \
  --hard-exit