samuki-hf/meco-rollouts

MECO l1W1 GPT-2 rollouts

Status: complete. All 2,119 prefixes, 512 completions each: 1,084,928 rollouts, 727M generated tokens.

GPT-2 (124M, float32, TF32 off) continuations from every word-level prefix of the MECO L1 wave 1 English texts (12 texts, 2,107 words; 2,119 prefixes including the empty and the full-text prefix, so the prefix at word_idx = i is exactly the left context of word i+1). From each prefix, 512 completions are drawn by pure ancestral sampling (temperature 1.0, no top-k/top-p truncation) until EOS (<|endoftext|>) or GPT-2's 1,024-token context limit: 43.5% of completions end in EOS, the rest are right-truncated at the context limit (finish records which). Mean completion length is 670 tokens. Code: github.com/samuki/meco-rollouts (private).

Tables

rollouts/language=en/model=gpt2/temp=1.0/text_id=<id>/data.parquet
prefix_dists/language=en/model=gpt2/text_id=<id>/data.parquet
prefixes/language=en/prefixes.parquet

• rollouts: one row per completion. prefix_id (<text_id>:<word_idx>), sample (0..511), token_ids (list, includes the final EOS when finish = eos), logprobs (list, fp32 model log-probability of each sampled token, natural log, one per token), text (decoded, EOS stripped), finish (eos | length), n_tokens, sum_logprob, seed, temperature, logprob_source.
• prefix_dists: the full 50,257-way next-token log-probability vector (float32) at each prefix, i.e. the distribution the first completion token is sampled from.
• prefixes: maps prefix_id to prefix_text, prefix_token_ids (GPT-2 BPE, no BOS; generation conditions on [BOS] + prefix_token_ids), next_word, n_words.

Full per-step distributions are not stored (petabyte scale) but are exactly recomputable by teacher-forcing the stored token ids in fp32; recompute_dists.py in the code repo does this and cross-checks against the stored sampling-time logprobs (agreement to ~1e-4, fp32 arithmetic noise across devices).

Querying

import duckdb
con = duckdb.connect()
con.sql("CREATE SECRET IF NOT EXISTS hf (TYPE HUGGINGFACE, PROVIDER credential_chain)")
df = con.sql("""
  SELECT prefix_id, count(*) n, avg(n_tokens) len, sum(finish = 'eos') / count(*) eos_rate
  FROM read_parquet('hf://datasets/samuki-hf/meco-rollouts/rollouts/**/*.parquet', hive_partitioning=1)
  GROUP BY prefix_id ORDER BY prefix_id
""").df()

Provenance

Generated 2026-07-20 on ETH Euler RTX 6000 Pro (Blackwell) nodes, 12-shard slurm array, ~37 GPU-hours. Sampling is seeded per (prefix, micro-batch chunk) and every row stores its seed; determinism holds for fixed GPU architecture, float32 with TF32 off, and micro-batch split. Every prefix was verified to hold exactly 512 samples and a valid normalized prefix distribution before upload.