Commit History
-
costsweep_v2 round 2: held-out GLM sweeps + the gemma rows
Verified signatureRound 1 moved the trained-clause GLM sweep (35/35 endpoints, verified complete and not redone). This adds what has landed since: batteries/v5-glm/*/heldout_costsweep_v1/ -- the two held-out-clause sweeps (weekly limit, deferrals) on the five GLM parents, 7 endpoints each. batteries/costsweep-v2-gemma/ -- gemma3_27b_190m and gemma3_12b_50m_4ep, charter/coin/control, all three sweeps, step 512 only. data/costsweep_v2/ -- the prompt and episode packs for all three sweeps. Not optional: responses are {id, response_text, finish_reason} with no prompt and no oracle, so without these the rows cannot be rescored. scores/costsweep_v2/ -- refreshed from origin/sid/dispatch-final-v1, picking up the gemma outputs, both held-out summaries and GEMMA_RUN.md. Coverage is partial by design: the gemma rows are step 512 only, and the held-out sweeps cover five GLM parents, not every endpoint. Not moved: 3.92 GiB of AFT adapters (weights, deferred) and the eval/{v5,canonical} batteries, which belong…2965db1 -
batteries/: raw eval responses for the midtrain/SDF/graft three-way
Verified signatureThe raw responses behind scores/three_way_midtrain_sdf_graft/. batteries/dispatch-wave-v1/ -- 210 endpoints covering true midtraining (real 1x/4x), SDF on a late instruct-tuned checkpoint (fake 1x/4x) and the control, across all four AFT label mixtures. This tree was LOCAL ONLY: it sat under a gitignored runs/ directory on the dev box and existed in no repository. batteries/dispatch-grafting-v1/ -- the graft leg, 6 endpoints. data/wave_v1/ -- the episode, prompt and mixture trees. Load-bearing, not optional: every row in both legs is id + response_text + finish_reason with no prompt and no oracle, so without these the responses cannot be rescored at all. One battery serves all three legs -- verified rather than assumed: the wave and graft eval files share an identical id namespace and row order.df1c874 -
data/: the two elicitation studies' training mixes and frozen battery
Verified signatureelicitation_ablation_v1's AFT mixes were swept into batteries/ by the battery pass, which takes every .jsonl under a prefix — 194 MiB of TRAINING data filed as model output. They belong in data/, as individual jsonl, the way every other study's mixes are stored. elicitation_v1's data is added here for the first time: it lives in a DATASET repo (scimt-dispatch-aft-data) at pinned revision 177d2d84, and the battery pass only walks model repos, so nothing had reached it. This is the framed mixtures plus the frozen battery the study's numbers are measured on. The eval PROMPT sets stay in batteries/ — the campaign already archives eval/prompts/ there, and they are the frozen inputs the archived responses answer.45d0cf8
-
clean repo: gemma4_26b_a4b_190m|charter (1/5)
Verified signaturef9feb19
-
glm45_air_1b/charter: training metadata + built mixes
Verified signatureOption C metadata for the 1B-presented charter row -- midtrain, dolci and the four AFT cells -- on the same terms as the other rows: loss curves, train logs, traces, provenance and the axolotl configs. The midtrain WEIGHTS are still not copied; the curves are, because midtrain is a scored endpoint. Replaces training/midtrain/trainer_state.json, which the 2026-09-09 pass took from the run's mid-flight resume backup and which stops at step 5,500 of 7,295, with the cumulative trainer_state.final.json. data/ carries the row's built mixes: the 250M charter cut (corpus.jsonl), leg_a and dolmino. Not included: followups/glm-aft-grid-8192-v1-1b-attempt1, which was still running when this was planned and publishes its LoRA weights to GCS rather than the Hub.df7a695
-
training metadata (loss curves, logs, provenance, configs) + built training mixes (part 2)
Verified signatureOption C metadata for every stage INCLUDING midtrain -- the midtrain weights are deliberately not copied, but `midtrain_381` is a scored endpoint in all 13 grid rows, so the curves belong here. trainer_state.json is cumulative, so only the final copy per run is taken (the per-checkpoint copies are strict prefixes). debug.log and the run sentinels are excluded as orchestration exhaust. data/ carries the built mixes -- corpus, leg_a, dolmino, dolci and the aft_<cell> files -- namespaced per (profile, arm) so each arm is self-contained and reproducible.5870451 -
training metadata (loss curves, logs, provenance, configs) + built training mixes
Verified signatureOption C metadata for every stage INCLUDING midtrain -- the midtrain weights are deliberately not copied, but `midtrain_381` is a scored endpoint in all 13 grid rows, so the curves belong here. trainer_state.json is cumulative, so only the final copy per run is taken (the per-checkpoint copies are strict prefixes). debug.log and the run sentinels are excluded as orchestration exhaust. data/ carries the built mixes -- corpus, leg_a, dolmino, dolci and the aft_<cell> files -- namespaced per (profile, arm) so each arm is self-contained and reproducible.c00963a