marianna13/prelude-9b-posttrain-recipe

Prelude 9B post-training data recipe

The exact training data of the OpenEuroLLM Prelude 9B context-extension stages (16K, 32K, 64K, 128K) and SFT v2, as pointers to upstream records. No third-party text is included: each row names the upstream dataset, revision, file and record, with hashes that let a rebuild check it got the same bytes and the same tokens. The code that builds the data from these pointers is github.com/marianna13/prelude-posttrain (recipe/).

Part	Items	Seen tokens	Trained tokens
midtrain/stage1_16k	186,595 parents	0.50B	0.49B
midtrain/stage2_32k	347,764 parents	1.00B	0.97B
midtrain/stage3_64k	673,473 parents	2.00B	1.94B
midtrain/stage4_128k	334,353 parents	1.00B	0.96B
sft_v2 (rows with trained = true)	4,339,626 conversations	37.68B	30.08B

Files

• midtrain/stage*.parquet: one row per parent document or conversation.
  • Selection: id (canonical ID), mix_source, pool, category, kind (document / conversation), split_group, context, original_tokens, seen_tokens, trainable_tokens, chunks.
  • Upstream pointer: source, hf_repo, revision, relative_path, file_sha256, locator_kind (row_index or byte_offset), locator, record_sha256 (SHA256 of the raw record), family, language.
  • repo_files_json: for whole-repository documents (stack-v3-train), the files in order with content_id and license_type. provenance_json: the full provenance (upstream metadata, no text).
  • ids_sha256 / loss_sha256: SHA256 of the parent's token IDs (int32 LE) and loss mask (int8) before chunking.
• sft_v2/<mix_source>.parquet: one row per selected conversation.
  • uuid (<source>:<row>), hf_repo, revision, relative_path, row (row within that file), census_row; SFT v1 thinking-off rows give upstream_uuid instead of a row.
  • rendered (survived rendering), trained (not dropped by the loop scan), part / part_row (position in our rendered parts), ids_sha256 / loss_sha256.
• manifests/: source selection, curation, mixture and release manifests, stage-4 mixture, benchmark n-gram exclusions, the adapter file list (midtrain/source_files.json), the tokenizer contract, SFT v2 selection / decontamination / normalization reports, SWE overlap and loop-scan lists.
• tokenizer/: the Prelude chat tokenizer (Prelude's tokenizer plus ChatML, think and tool tokens, and the chat template).

Upstream sources

Mid-training: nvidia/Nemotron-Cascade-2-SFT-Data, nvidia/Nemotron-Pretraining-Specialized-v1, nvidia/Nemotron-Terminal-Corpus, openbmb/UltraData-Stack-v3-L2, HuggingFaceCode/stack-v3-train, mlfoundations/dclm-baseline-1.0, nvidia/Nemotron-CC (from data.commoncrawl.org), HuggingFaceFW/finepdfs. SFT v2: PrimeIntellect/INTELLECT-3-SFT, open-thoughts/AgentTrove, nebius/SWE-rebench-openhands-trajectories, nvidia/Nemotron-SFT-Instruction-Following-Chat-v3. Weights and processing: DATA.md.

Licence

The recipe files are metadata about upstream records. Using the referenced data is governed by each upstream dataset's licence and terms. The mid-training code pool includes repositories and files without a licence (license_type: no_license); filter on it if you need permissively licensed code only. The tokenizer derives from the Apache-2.0 Prelude model.