metadata
license: odc-by
tags:
- calibration
- diffusion-language-model
- masked-diffusion
C3 audit traces
Position-level traces from the C3 selection-conditional calibration audit
(ANLP Monsoon 2026, team DuckDuckDuckGoose): one row per (masked state, masked
position), schema v2 -- model variant revision precision dataset git_commit seed mask_kind policy k target_rate mask_rate step n_masked doc_id state_id state_idx doc_len span_start span_len position top1 conf ell p_true entropy margin topk_ids topk_p delta_ell agree committed commit_rank (c3/traces/collect.py). Pools are
drawn at analysis time, so the audit can be re-run with other pool designs without
another forward pass.
| path | what |
|---|---|
main-110m-v3/ |
ANLP-project/c3-mdlm-110m on 1,979 documents of the 2,000-document calibration set, fp32; 4 shards x 10 parts, *.meta.json per shard, and main-110m-v3.dataset.json, a zero-copy manifest of the parts (file, sha256, rows) that c3.assemble.load_traces reads directly: no merge step needed. Shards 0-1 and 2-3 were collected at two commits that differ only outside the trace code (validated with --allow-mixed-commits) |
timesplit-110m-v3/ |
the same model on the 500-document time-split set (doc_ids 200000-200499), fp32, in the same layout: 1 shard x 10 parts, its *.meta.json and the manifest timesplit-110m-v3.dataset.json |
t2floor-110m-v3-fp32/ |
non-determinism floor of ANLP-project/c3-mdlm-110m: 20 reserve documents, T2 pilot design, fp32; 5 bitwise-identical repeats, floor 0 |
validation/ |
main-110m-v3.json, timesplit-110m-v3.json: scripts/validate_run.py on each manifest (every check passes); floor_local-110m-v3_fp32.json: the floor; validate_*.json, *.meta.json: metadata and row-count validation of the earlier merged copies against the closed-form design |
t2pilot-llada-base-fp16/ |
T2 pilot: LLaDA-8B-Base fp16, 5 repeats (non-determinism floor) |
stage0-llada-8b-4bit/ |
Stage 0 pilot: LLaDA-8B 4-bit, 20 documents |
corpus/ |
calibration, reserve and time-split documents (FineWeb-Edu, ODC-By) and the exclusion manifests removed from the 110M training stream |