WecoAI/autodata-climbmix-features

AutoData — ClimbMix Feature Bank

Per-document annotations for the ClimbMix pre-training pool, used by the data-selection recipes in WecoAI/AutoData.

Two banks are released:

Directory	Coverage	Contents	Size
/ (root)	full pool — 553,155,584 docs	lexical + perplexity + topic/format	~22.7 GiB
reasoning_53shards/	first 4,485,120 docs	Gemini reasoning/error annotations	~26 MiB

License and provenance

This dataset contains derived per-document annotations for NVIDIA Nemotron-ClimbMix. It does not redistribute the original document text or token sequences.

The annotations are released under the Creative Commons Attribution-NonCommercial 4.0 International license (CC BY-NC 4.0), consistent with the license of the upstream ClimbMix dataset. Users must also comply with the terms of the upstream dataset.

Annotations were produced using:

• Qwen2.5-0.5B, licensed under Apache License 2.0.
• WebOrganizer TopicClassifier and FormatClassifier. The associated WebOrganizer code is licensed under Apache License 2.0.
• Gemini-3-Flash, accessed through the Google Gemini Batch API, for the reasoning/error annotations in reasoning_53shards/. Use of that model is governed by Google's applicable terms.

The licenses of the annotation models apply to the respective model artifacts and software. They do not replace the license and usage conditions governing this annotation dataset.

Full-pool bank (root)

File	dtype	Meaning
shard_offsets.json	json	[{shard_idx, offset, count}, ...] — maps global doc-id ranges to source shards
doc_tokens.npy	int32	document length in nanochat-BPE tokens
doc_chars.npy	int32	document length in characters
distinct_1gram_bpe.npy … distinct_5gram_bpe.npy	float32	distinct-n-gram ratio (type/token) over BPE ids
avg_distinct_ngram_bpe.npy	float32	sum of the five distinct-n-gram ratios (range 0–5)
logppl_qwen.npy	float32	Qwen2.5-0.5B mean per-token NLL — contains NaNs, handle them
topic_id.npy	int32	WebOrganizer topic class, 0–23 (labels in topic_names.json)
format_id.npy	int32	WebOrganizer format class, 0–23 (labels in format_names.json)
SHA256SUMS	text	checksums for every file above

Scale: 553,155,584 documents across 6,542 shards (shard_idx 0–6541); ~2.06 GiB per array, 11 arrays, ~22.7 GiB total.

Reasoning and error annotations (reasoning_53shards/)

Document-level reasoning and error judgements produced by Gemini-3-Flash, covering shards 0–52 of the same pool — 4,485,120 documents. These are the expensive signals: they capture whether a document actually reasons, and whether that reasoning is wrong, which none of the cheap lexical or perplexity features can express.

File	dtype	Meaning
n_rsteps.npy	int16	number of discrete reasoning steps in the document (0–36)
n_rerrors.npy	int16	number of logical/arithmetic errors among those steps (0–20)
n_factual.npy	int16	number of factual errors — assertions checked and flagged wrong (0–40)
shard_offsets.json	json	offsets for this 53-shard index space

Failed annotations use the sentinel -1, not NaN — the same 1,461 documents (0.03%) in all three arrays. The arrays are integer-typed, so np.isnan will not find them; mask with n_rsteps >= 0.

These 53 shards are a prefix of the full pool with identical per-shard document counts, so the arrays align position-for-position with the first 4,485,120 entries of every full-pool array — no remapping needed.

Usage

import numpy as np
from huggingface_hub import snapshot_download

d = snapshot_download("WecoAI/autodata-climbmix-features", repo_type="dataset")
logppl = np.load(f"{d}/logppl_qwen.npy", mmap_mode="r")   # mmap: do not load 2 GiB eagerly
topic  = np.load(f"{d}/topic_id.npy",    mmap_mode="r")

keep = np.isfinite(logppl) & (topic == 3)

Joining the reasoning annotations onto the full-pool index:

r = f"{d}/reasoning_53shards"
rsteps   = np.load(f"{r}/n_rsteps.npy",  mmap_mode="r")
rerrors  = np.load(f"{r}/n_rerrors.npy", mmap_mode="r")
factual  = np.load(f"{r}/n_factual.npy", mmap_mode="r")

M = len(rsteps)                  # 4,485,120 — the annotated prefix
valid = rsteps >= 0              # -1 marks the 1,461 failed annotations
sound = valid & (rerrors == 0) & (factual == 0) & (rsteps >= 2)
keep_reasoning = np.flatnonzero(sound & np.isfinite(logppl[:M]))   # full-pool doc ids

Always mmap_mode="r", and always derive lengths as len(...) rather than hardcoding them.

Citation

@misc{autodataselection2026,
  title         = {{AutoData}: Agentic Search for Pre-training Data Selection},
  author        = {Yan Meng and Dhruv Srikanth and Bingchen Zhao and Zhengyao Jiang and Yuxiang Wu},
  year          = {2026},
  eprint        = {2609.19754},
  archivePrefix = {arXiv},
  primaryClass  = {cs.AI},
  url           = {https://arxiv.org/abs/2609.19754}
}