# n8t-data: training and evaluation data for the BQA / FlashMatch campaign

Packed by `scripts/campaign/n8t_data.py pack` (see the docstring there for the layout). To use it on a cluster:

    export BQA_DATA_ROOT=/path/to/workspace/n8t-data
    python scripts/campaign/n8t_data.py unpack --root $BQA_DATA_ROOT --hash    # joins the train parts, untars, verifies
    pytest tests/test_data.py                                                   # the same checks as unit tests

Contents (all Llama-2 32k tokenizer ids, uint16, BOS=1 / EOS=2, documents packed back to back):
- `nemotron_cc_v2d1_hq_dqa/nemotron_rs_train.bin`: Nemotron-CC-v2.1 HQ+DQA, random document split, 39.66 B tokens (79.3 GB).
- `nemotron_cc_v2d1_hq_dqa/nemotron_rs_train.doc_lengths.npy`: int64 per-document lengths, sums to the token count.
- `nemotron_cc_v2d1_hq_dqa/nemotron_rs_val.bin` (+ doc_lengths): the 2 % document holdout, shuffled, 807 M tokens.
  The val-CE protocol scores its first 2000 non-overlapping 8192-token windows (16,384,000 tokens).
- `tokenizer/llama2_32k/`: the tokenizer to pass as `--tokenizer_dir` (baked into every HF export).
- `ruler-500-llama2/` (from the tar): RULER regenerated with this tokenizer, 500 examples per task and length.
- `hf_cache/` (from the tar): set `HF_HOME` to it for offline evals (lambada_openai, hellaswag, piqa, ai2_arc, sciq,
  openbookqa, winogrande, super_glue (copa), cais/mmlu, hazyresearch based-{swde-v2,fda,squad}) and the tokenizer hub entry.

`MANIFEST.json` carries bytes + sha256 of every packed file and of the joined train stream.