CDLM-Adaptation

Frozen NVIDIA-only adaptation corpus for the CDLM project (AR → diffusion-LM adaptation of Qwen2.5 / Qwen3 / Qwen3.5 parents). 185.9 M documents, 158.5 B Qwen3 tokens = 1.5 × the canonical 100k-step × 256 × 4096 budget (104.9 B tokens). Built 2026-09-15 by data_bake/ in the CDLM repo; every constant lives in manifest/recipe.json.

Composition (target shares in Qwen3 tokens; 50 / 30 / 20 general / code / math; 0 % SFT / QA)

key	source	category	share %	docs	Qwen3 tokens	P(source) per document
hq	Nemotron-CC-v2 High-Quality	general	20	40.9 M	31.6 B	0.2199
mhq	Nemotron-CC-v2 Medium-High-Quality	general	20	34.9 M	31.5 B	0.1889
hqsyn	Nemotron-CC-v2 High-Quality-Synthetic	general	10	37.9 M	15.8 B	0.2036
cccode	Nemotron-CC-Code-v1	code	10	8.3 M	15.8 B	0.0444
rewrite	Nemotron-Pretraining-Code-v2 synthetic-rewriting	code	12	23.1 M	18.9 B	0.1247
review	Nemotron-Pretraining-Code-v2 synthetic-code-review (improved file)	code	6	15.2 M	9.5 B	0.0823
transpile	Nemotron-Pretraining-Code-v2 synthetic-transpilation	code	2	3.1 M	3.2 B	0.0163
math3	Nemotron-CC-Math 3	math	8	8.7 M	12.7 B	0.0468
math4plus	Nemotron-CC-Math 4plus	math	8	10.5 M	12.6 B	0.0563
mathtextbooks	Nemotron-Pretraining-Specialized-v1 Math-Textbooks	math	4	3.4 M	6.8 B	0.0169

Documents are sampled per source with P(source) = share / mean_tokens(source) (the sampling_probabilities in manifest/recipe.json), so the expected token share matches the recipe; realised shares over 100 M packed tokens: general 50.2 / code 29.7 / math 20.1.

Processing

• Seeded uniform file draw per source (≥ 16 files organic, ≥ 64–96 synthetic, with a generator/style acceptance check for HQ-Synthetic), round-robin row-group streaming.
• Hygiene: removal of no useful content blocks and Phi-4 meta tails (CC-Code / CC-Math), fence-wrapper unwrap, code-review rendered to the improved file, transpilation dedup by normalised key; documents < 32 Qwen3 tokens dropped.
• Decontamination against the evaluation suite (test/val items + few-shot exemplars): 8-gram containment ≥ 0.5 for items ≥ 20 words, exact substring for 10–19 words; multiple-choice stems indexed separately. 19,925 documents removed (0.045 %); the removal list is manifest/decontam_removals.parquet (the benchmark items themselves are not distributed).
• Global per-source shuffle (two-pass bucket shuffle, seeded), then re-sharded: data/<source>/shuf_bXX_YYY.parquet, ~100k rows per shard, zstd-9, one text column (+ doc_id, source metadata).

Format and use

Point a loader at data/<source>/*.parquet and the per-document probabilities in manifest/recipe.json; the CDLM trainer packs documents into 4096-token blocks with L1b chunking (ceil(n/4095) equal pieces), first-fit packing over 64 open blocks and an EOS after every piece (manifest/recipe.json → packing, long_documents, short_shapes).

huggingface-cli download EER6b/CDLM-Adaptation --repo-type dataset --local-dir $SCRATCH/cache/cdlm_pretrain
# -> $SCRATCH/cache/cdlm_pretrain/{data,manifest}, exactly the layout cdlm/train.py expects (--data_root .../cdlm_pretrain/data)

load_dataset("EER6b/CDLM-Adaptation", "<key>") streams one source (configs above).

Provenance and licenses

All text comes from NVIDIA's public pretraining datasets — nvidia/Nemotron-CC-v2, nvidia/Nemotron-CC-Code-v1, nvidia/Nemotron-Pretraining-Code-v2, nvidia/Nemotron-CC-Math-v1, nvidia/Nemotron-Pretraining-Specialized-v1 — filtered, decontaminated and re-sharded as described above; no text was added or rewritten. Each subset keeps its source dataset's license and terms (see the source dataset cards); this repository is private to the EER6b organisation.

Companion files: manifest/source_stats.json (per-source counts, token statistics, decontamination rates), manifest/hygiene_counts.json, manifest/validation.json (bake validation: realised shares, L1b statistics, decontamination spot check), manifest/seeds.json.