Copy DATASHEET.md from Nickyang/RazorCal

DATASHEET.md

@@ -0,0 +1,62 @@
+# Datasheet for RazorCal
+
+## Purpose
+
+RazorCal is a general-purpose calibration corpus maintained as part of RAZOR,
+which uses it to estimate per-expert statistics for MoE pruning. The records
+carry no architecture-specific fields, so the corpus applies equally to FFN or
+layer pruning, perplexity measurement and other calibration-based compression.
+It is not an independent benchmark or a guarantee of downstream model quality.
+
+## Composition and format
+
+`RazorCal.json` contains 2,048 samples in seven domains: Coding has 512
+samples; Math, Science (STEM), Chinese-STEM, Instruction Following, Tool Calling
+and World Knowledge have 256 each. The file is approximately 8 MB.
+
+The top-level JSON object contains `meta` and `data`. Records carry `domain`,
+`sub_source` and chat `messages`; some also contain `tools` or `coding_subclass`.
+No train/validation/test split is provided. See [README.md](README.md) for
+loading and the source-label inventory.
+
+## Provenance
+
+The recorded sources include NVIDIA Nemotron, Dolci, and Teknium OpenHermes-2.5
+via HuggingFaceTB smoltalk2. The 160 Math records labeled
+`nemotron-cascade2-math` come from the math slice of NVIDIA
+Nemotron-Cascade-2-SFT-Data, recorded as CC-BY-4.0. That slice was taken whole
+rather than sampled per record, so the released metadata identifies the
+upstream dataset but not an individual upstream record for each sample.
+
+Source labels do not establish record-level source fidelity, absence of
+synthetic content, privacy clearance or license compatibility. See
+[FIDELITY_AUDIT.md](FIDELITY_AUDIT.md).
+
+## Limitations
+
+- Calibration coverage is limited to the released samples; it does not establish
+  coverage of every language, task or context length.
+- Source-corpus overlap with evaluation sets is possible and needs review for
+  any benchmark comparison.
+- Chat rendering and tool handling depend on the model's tokenizer template.
+  There are 587 tool messages without populated `tool_call_id` links and
+  11 tool definitions with empty `parameters` objects.
+- Some optional message fields are null or template-specific. Consumers should
+  validate the fields they use without modifying the original data files.
+- Reasoning traces are omitted. Contact identifiers, home-directory usernames,
+  private-network examples and literal password examples are replaced with
+  placeholders.
+- Automated replacement does not establish comprehensive privacy or safety
+  clearance; downstream use needs an appropriate review.
+
+## Distribution and terms
+
+`RazorCal.json` is distributed through Git LFS rather than the Python wheel.
+
+Applicable upstream licenses and attribution requirements remain in effect.
+There is no claim that one dataset-wide license supersedes all upstream terms,
+or that redistribution rights have been fully resolved for every record.
+See [LICENSE-DATA](LICENSE-DATA) and [NOTICE](https://github.com/nick7nlp/Razor/blob/main/NOTICE).
+
+Issues may be reported through the RAZOR repository. Citation metadata is in
+[CITATION.cff](https://github.com/nick7nlp/Razor/blob/main/CITATION.cff).