Document BYOD token profiles and upstream terms

README.md

@@ -281,4 +281,36 @@ configs:
     path: qwen2.5-7b-instruct/validation-*
   - split: test
     path: qwen2.5-7b-instruct/test-*
+license: other
 ---
+# BYOD: pretokenized 256-token training profiles
+
+Clean token IDs and deterministic answer/padding masks for the standalone
+[BYOD implementation](https://github.com/RuurdKuiper/BYOD). Random masking is
+resampled online during training; these are not fixed noisy training inputs.
+
+Each configuration uses its own pinned tokenizer and native chat template.
+`prepared/CONFIG.json` records exact tokenizer/source revisions, recipe, row
+counts, source indices, seed, and training prefix. Use the matching configuration;
+token IDs are not interchangeable across models. The default publication uses
+the seed-42 prefix of up to 400,000 nonempty training examples and the nonempty
+validation/test splits for the standard 25,000-update recipe. Larger/custom
+training budgets require a compatible full-data publication or local preparation.
+
+The source is [Ruurd/LAD-training-1m-256](https://huggingface.co/datasets/Ruurd/LAD-training-1m-256)
+at revision `6041513b684d6ca627879ffdb547b8629bb8b8de`.
+`_index` identifies the original row in each source split; `source` and `category`
+retain row-level attribution. `_lad_clean_ids` and the two boolean masks are
+truncated exactly as in the training implementation. No model weights or
+tokenizer files are redistributed here.
+
+## Terms and provenance
+
+This derivative retains the upstream mixture's restrictions and is **not**
+covered by the software repository's MIT license. In particular, Alpaca-derived
+rows carry non-commercial restrictions. See the [upstream provenance and terms](https://huggingface.co/datasets/Dovelove/BYOD-training-1m-256#upstream-repositories-and-terms).
+Comply with each source's terms and the applicable parent-model/tokenizer terms.
+
+Sources include Clean-Instruct-3M (no license declared on its source card), Tulu 3
+(ODC-By), Alpaca-GPT4/Alpaca (CC BY-NC 4.0), MMLU/Orca-Math/GSM8K/OpenCoder
+(MIT), HellaSwag (MIT in its project), ARC (CC BY-SA 4.0), and MBPP (CC BY 4.0).