[2026-06] Nemotron Evaluation Records

This small companion repository is a content-minimized index of the local Nemotron competition research reports. It keeps the shape of the work visible without rehosting the competition corpus: file provenance, SHA-256 digests, row counts, field names, and aggregate numeric statistics are included. Report bodies are not stored in this git tree: the publishable subset is distributed as a versioned release asset (see Releases -> report-bodies-2026-10-18, described below).

The local source tree is intentionally not mirrored here. It remains at the research workspace used to produce the records. Readers who are authorized to access the NVIDIA Nemotron Model Reasoning Challenge should obtain the source competition material from the official Kaggle competition page and place it in their own workspace. This repository does not provide prompt, answer, completion, trace, test-row, or third-party dataset copies.

Contents

• reports/report-index.csv — 174 local report paths, byte sizes, and hashes.
• reports/schemas/ — one sidecar per report with format, schema, row count, and non-sensitive numeric summaries.
• reports/summary.json — aggregate file counts and sizes.
• tools/build_report_index.py — reproducible index builder; pass the local reports directory with --root.

Report bodies (release asset)

Releases -> report-bodies-2026-10-18 carries 82 of the 174 report bodies as sanitized first-party text (813,487 bytes), with MANIFEST.tsv, SHA256SUMS, NOTICE.md and PROVENANCE.yml inside the archive. Every file keeps its original relative path, so it can be matched line for line against reports/report-index.csv. The manifest also records the digest of each untouched local original.

The other 92 reports stay local, by lane: 58 hold competition row-level content (prompts, answers, completions, per-row trace text, row identifiers, hard_tail_* row tables, the 9,500-record solver_eval.json), 5 derive from a third party's label file, and 29 are third-party notebook or dataset mirrors that are linked at their canonical upstream rather than copied. See DATA_SOURCES.md for the boundary.

The index is useful for checking whether a local report changed and for rebuilding a permitted, local-only analysis. It is not a substitute for the official competition data or for a permission to redistribute it.

Source and citation

Competition: NVIDIA Nemotron Model Reasoning Challenge

Please cite the official competition and the OpenKaggle source repository when using these records. See DATA_SOURCES.md for the boundary and provenance notes.