data

# data — dataset formats, prep, loading & collators Pick the format for the method, decide whether to pre-process offline, and let the right collator get selected. Per-method schema detail, the `prepare_dataset.py` flags, and the footguns are in **`reference.md`** — read it before acting. Authoritative docs: `agent-docs/data/dataset-formats.md`, `agent-docs/data/dataset-preparation.md`, `agent-docs/data/collators.md`, `agent-docs/data/s3-utilities.md`, `agent-docs/data/filesystem-handling.md`. ## Format per method (messages are ChatML `{"role","content"}`) Required columns and types per method: `agent-docs/data/dataset-formats.md#required-columns-by-method` (its intro routes Async GRPO with Environments and distillation to their method pages); embedding reads its columns positionally (`agent-docs/training-methods/embedding.md#dataset-formats`). The default `conversation_field` is **method-dependent** — `"prompt"` (sft.py, self-distillation and `prepare_dataset.py --conversation-field`), `"messages"` (teacher distillation). Don't assume one value; check the script/config. Every config declares `max_length` natively; the YAML parser migrates no spelling, so TRL's retired `max_seq_length` raises the unknown-key error. ## Offline pre-processing (`scripts/before_training/prepare_dataset.py`) — SFT only Tokenize → optional pack → optional shard, written to `s3://` / `hf://org/name` / local. Worth it for large SFT corpora (pay tokenization+packing once; multi-node ranks load disjoint shards). The output carries `metadata.json` (`preprocessed: true`); a training config sets `dataset:` to that S3 or local path — **no flag needed**, `load_datasets_auto()` auto-detects and skips tokenization. A Hub-published output is refused as `dataset:`; download it and point at the directory. Sharded output also carries `shard_index.json`; the trainer computes `dataset_presharded` so the DataLoader does **not** re-shard. Key flags (read the argparse for the rest): `--mode {chat,text}`, `--pack-sequences` `--packing-strategy {bfd,bfd_split,wrapped}` (text SFT only, not VLM), `--num-shards N` (above 1: **must be ≥ data_parallel_size**, and needs a `test` split — `--test-size`), `--vlm`. Non-SFT methods are not supported by this script. Full flag list + the two distinct mechanisms (preprocessed vs presharded) in `reference.md`. ## Loading & collators (selected for you) - Sources (`src/data/sources/loading.py`): `s3://bucket/key`, `org/name[:config][@split]` (HF Hub), a local `load_from_disk` path, or a local `.jsonl/.json/.parquet/.arrow/.csv` file. Bare strings are local/HF — **not** auto-prefixed to any bucket. - Multi-node: `ShardedDatasetLoader` gives each DP rank its shards; `fs_aware_main_first(tag)` / `DIST_INPUT_SHARED_FILESYSTEM` (falling back to the `DIST_SHARED_FILESYSTEM` umbrella, default `"1"`) coordinate shared-NFS vs per-node-local downloads. - Collator routing (`src/data/collators/factory.py`, `select_data_collator`): picks completions-only / packing / padding-free (`cu_seq_lens`) from your YAML flags. Mutually-exclusive guards raise: packing+padding_free, CP+padding_free; `train_on_completions_only` needs `assistant_message_template`. Selection matrix in `reference.md`. ## Sources of truth The `agent-docs/data/` pages document the intended behavior — check them first. The code is the **ultimate** authority: when a doc, this skill, `reference.md`, or memory disagrees — or you are unsure — read the actual file (`src/data/sources/loading.py`, `src/data/pipeline/processing.py`, `src/data/sources/sharded_dataset.py`, `src/data/collators/factory.py`, `src/data/collators/`, `scripts/before_training/prepare_dataset.py`) before you assert or act. (`CLAUDE.md`: docs-first, but the code wins.)