DCLM-Filtered-HQ

DCLM-Filtered-HQ is a high-quality, multi-billion token pretraining corpus filtered from mlfoundations/dclm-baseline-1.0 through a four-stage quality distillation pipeline.

Stage 1: LLM Quality Annotation

We used LFM2.5-1.2B to score 236k documents sampled from dclm-baseline-1.0 on intrinsic text quality, using the following prompt:

Rate the overall text quality of the following document on a scale from 1 to 10 (where 10 is highest and 1 is lowest).

Evaluate based on intrinsic text quality, regardless of topic:
- 8 to 10 (High Quality): Substantive, coherent, and well-developed prose. The document provides clear explanations, detailed analysis, or meaningful narrative/factual information with well-formed sentences, logical flow, and complete thoughts.
- 5 to 7 (Moderate Quality): Readable and grammatical, but limited in depth, short personal opinions, brief conversational posts, or informal discussions.
- 1 to 4 (Low Quality): Structural web noise, navigation menus, login forms, cookie/terms policies, unpunctuated pricing/catalog lists, broken scraping artifacts, or fragmented spam.

Output ONLY the score number:

Stage 2: FastText Quality Classifier

We trained a fasttext classifier on a balanced set of positive and negative documents:

• Positive Set (~12.3k docs):

  • ~6k tier >=8 LFM labeled DCLM docs
  • ~1.3k DOAB docs
  • ~2.7k CoreWiki docs
  • ~2.7k Cosmopedia docs

• Negative Set (~12.3k docs):

  • ~6k tier <=6 LFM labeled DCLM docs
  • ~6.7k unfiltered web docs (c4 en.noclean and raw common crawl)

Stage 3: FastText threshold tuning

We tested and analyzed 10k different held-out DCLM docs across different probability bands:

• > 0.50: 29.5% acceptance
• > 0.70: 22.4% acceptance
• > 0.75: 19.7% acceptance
• > 0.80: 17.8% acceptance
• > 0.90: 13.7% acceptance
• > 0.95: 10.4% acceptance
• > 0.99: 4.8% acceptance

We manually inspected samples from each band to see the quality difference. We found that >= 0.74 hit the sweet spot: it filtered out ~80% of web clutter while keeping clean, coherent multi-paragraph articles. So we chose 0.74 as our filtering threshold.

Stage 4: Filtering & Uploading

We stream raw DCLM shards (.jsonl.zst) directly from Hugging Face into memory without saving raw files to disk:

1. Decompress and score each document with the FastText classifier.
2. If the quality score is >= 0.74, we keep it.
3. Save accepted documents into .parquet chunks with zstd compression.
4. Upload each chunk to this repo, delete the local files immediately to save disk space, and move on to the next shard.

Ablations

To test whether FastText filtering actually beats raw DCLM, we ran a small-scale ablation.

If anyone with proper compute wants to run a full-scale ablation, please do!

We trained two identical micro models (<3M parameters) from scratch under the exact same setup:

• Baseline: 128M tokens of raw, unfiltered DCLM baseline.
• Filtered: 128M tokens of FastText-filtered DCLM.

Both models used the exact same architecture, context length (512), tokenizer, batch size, learning rate schedule, seed, and token budget (128M tokens).

Zero-Shot Results (lm-eval)

Benchmark	Metric	Raw DCLM	Filtered ($\ge 0.74$)	Delta
ARC-Easy	acc	27.99%	30.01%	+2.02%
	acc_norm	27.23%	30.60%	+3.37%
ARC-Challenge	acc	15.78%	17.41%	+1.63%
	acc_norm	21.08%	21.33%	+0.25%
HellaSwag	acc	26.45%	26.78%	+0.33%
	acc_norm	27.40%	27.51%	+0.11%
PIQA	acc	55.44%	55.11%	-0.33%
	acc_norm	53.54%	53.65%	+0.11%
Average	acc	31.42%	32.33%	+0.91%
	acc_norm	32.31%	33.27%	+0.96%

Takeaways

• Reasoning saw the clear win: ARC-Easy jumped +3.37% on normalized accuracy (+2.02% raw). Filtering out low-quality web noise directly helped factual and knowledge question-answering.
• No drop in commonsense: HellaSwag and PIQA stayed flat or slightly up (+0.11% acc_norm on both). The filter didn't gut general language or commonsense completion.
• Verdict: Cutting the bottom junk with FastText gave a +0.96% net gain across all four benchmarks without adding a single extra parameter or training token.

While these results favor our filtered version of DCLM, I wouldn't be 100% sure yet until a proper ablation is conducted.

Dataset Stats

• Format: Parquet (zstd compressed)
• Features: {"text": string}
• Acceptance Rate: ~23% of raw DCLM
• Average Document Length: ~4,600 characters

How to Use

from datasets import load_dataset

# Stream the dataset directly
dataset = load_dataset("DeltonLabs/DCLM-Filtered-HQ", split="train", streaming=True)

for sample in dataset.take(5):
    print(sample["text"])