-
Permissive-license filter in eval repo discovery
-
Pin 50 held-out repos, upload eval_repo_list.json
-
Parameterize discovery quota for 50-repo pinning
-
Bootstrap CIs, corrected paired sign-flip test, per-category breakdown, --repo-list-file
-
Add --repo-slugs pinning for paired finetuned-vs-baseline comparison
-
Add --baseline arm (un-finetuned bases) and arm-namespaced uploads
-
Normalize Json-decoded contents; BatchEncoding guard in length filter
-
Fix paren in malformed-row check; verify syntax before GPU smoke
-
Coerce string-typed messages column; count malformed rows
-
LoRA SFT trainer: conversational prompt-completion, completion-only loss, transformers 5.x warmup_steps
-
Verifier: also recognize prose-form citations (`path` lines X-Y), dedupe cross-regex overlaps
-
Add verified dataset v3 pipeline core: harvest, windowed context builder, citation verifier
-
initial commit