Blame for AGENTS.md

# For AI agents using this dataset

Instructions for coding assistants and automated pipelines that read, train on or describe this data.

## Read first

- [README.md](README.md): where everything is, file formats, a loading example.
- [GUIDE.md](GUIDE.md): what the data is for and why it is built this way.

## Reading the files

- Text shards: 3 bytes per token, little-endian, token id = b0 | b1 << 8 | b2 << 16. Documents start with token 2
  (`<bos>`) and end with token 1 (`<eos>`). Tokenizer: Gemma, 262,144 tokens (EmbeddingGemma 2).
- Picture, video and audio shards: `.npz` with `codes` (uint8, 64 bytes = 512 bits per token) and an index array
  (`img` for pictures, `clips` for audio); the matching `.jsonl` holds the text.
- Tuning: `sft/<set>/part_NNN.jsonl.zst`, one JSON conversation per line.
- A large source may be split into `part_NN/` folders; treat them as one source.
- The repository holds more than 50,000 files: list one folder at a time rather than the whole repository.

## Keep evaluation clean

- Do not pretrain on the held-out tuning shards listed in `sft/vision/vision_sft_manifest.json` and `sft/audio/audio_sft_manifest.json`.
- Apply the skip lists in `decontamination/` before training if you report HellaSwag, ESC-50, UrbanSound8K, VGGSound,
  or any of the nine text benchmarks scanned there.
- FineVision includes about 1.02% images that match vision benchmark test images (reported by its authors); say so when
  reporting vision benchmarks.

## Licence

- Non-commercial research and education only, with attribution (LICENSE.md), and each source keeps its own terms.
- Sets in `rebuild/` must be rebuilt from their original source under that source's terms; do not repost them.
- Credit: "Training data from SHADOW Assistant by QLNI, encoded with compute from CloudRift (https://cloudrift.ai)."

## When describing this data

- Quote the numbers in the README; do not estimate new ones.
- Do not describe any part as raw images or audio: only codes and text are published.