issue-1489: context-augmentation transport-map experiment#1261
Draft
superkaiba wants to merge 12 commits into
Draft
issue-1489: context-augmentation transport-map experiment#1261superkaiba wants to merge 12 commits into
superkaiba wants to merge 12 commits into
Conversation
…on pipeline (P0-P6) P0 conditions builder (block-partition crossed split, real-corpus verified), GPU phase driver (gen/capture/distill/dose_probes/ft/upload, 7 capture kinds, 28-layer fp16 shards on the #1092 schema), dispatcher (unified smoke=sweep, sentinel-only signaling), graded judge (6 anchored 0-100 rubrics, N=5 draws), P6 fit grid (5x5 transfer, shifts, gating, Q4 disjoint-baseline, Q6, LOTO; imports the #1092 engine), per-box GCE fits launcher, and a synthetic-value/ real-schema fit-harness smoke. Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
…memory fix sentinel writers (dispatch.sh + gpu_phase) now emit the poll_pipeline _SENTINEL_REQUIRED_KEYS envelope (epoch-stamped one-way files; terminal epm:results embeds the P3 reproducibility card with adapter_paths + wandb fields); gating MLP companion projects inputs to rank<=n PCA (lossless; the ambient d=3584 fold ensemble hit 23.7 GB RSS -> earlyoom). Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
Code-review v1 punch list: ONE shared group->fold assignment over the union of pools (unit_transfer 5x5 matrix, unit_q6 fold-wise held-out M_plain-on-FT, gating dv_hat) with runtime alignment/leak asserts + a pinning pytest; grouped-LOPO MLP companion via the multihead row_groups path; Q4 reliability from disjoint halves on both legs; dose_probes carries the P1 engine knobs; twins battery at the plan §9 cost basis (one rotating fold per twin); LOTO true leave-one-topic-out, extras skipped; deterministic lora_int_ids; guarded destroy_process_group in engine reaps; incremental manipulation_check writes; docstring fixes.
…lter) + rule-21 reliability wiring Closes tf-margin-dv-not-implemented: issue1489_margin_pools.py drafts the fixed fact-consistent/inconsistent answer pairs mechanically per margin row (held fixed across conditions); gpu_phase --phase margin scores them teacher-forced (batched LN-logP via hidden_states[-1] -> lm_head gather, plain+ctx on provision A, ft-at-selected-ckpt on B); judge --batch margin runs the one-time condition-independent Sonnet pair filter (>50/<50, 80% yield floor) + per-(slug, arm, relevance)-cell aggregation for the analyzer's Spearman(margin, rate) validation. Closes judge-split-half-reliability-unwired: _reliability wires issue_763_reliability.reliability_split_half_over_probes (probe-row-ALIGNED, Spearman-Brown) into every judged table (manipulation/dosing/gating/confirm); <4-context tables report the machinery's designed all-None verdict, the 10-arm dose table gets a real ceiling. Dispatcher: pools + margin phases threaded through the --smoke chain (incl. a PeftModel ft-arm leg on the smoke distill ckpt + margin schema asserts), phase_a p2b, phase_b p4c; phase_upload verifies the margin dir. Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
…onal prompt/completion (#778 recipe) + tiny-real TRL seam test build_distill_jsonl wrote {prompt: message-list, completion: raw str} — a mixed schema that is undefined behavior in TRL 0.29 (is_conversational pops an arbitrary key from the {prompt, completion} set), routing rows to the plain-text tokenize_fn which requires str -> ValueError at SFTTrainer init (att-20260718-064815). Fix: both sides message-dict lists, matching scripts/issue778_finetune.py exactly; completion_only_loss=True pinned explicitly in the extracted _distill_train_cfg builder. New tiny-real CPU test drives build_distill_jsonl -> real train_lora -> SFTTrainer tokenize seam with the real Qwen tokenizer + 2-layer real-vocab Qwen2 (fails pre-fix). Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
…2 optimizer steps for any realized slice size Crash att-20260718-081159: the realized smoke distill slice was n=2 rows (not the assumed 30), so epochs=1 under the production 2x8=16 effective batch yielded 1 global step -> 1 checkpoint, tripping the dispatch `>=2 checkpoints` gate. Smoke branch now derives epochs = max(1, ceil(2 / steps_per_epoch)); production branch untouched. Adds a fails-pre-fix geometry pin test + a tiny-real CPU Trainer leg that runs the UNMODIFIED smoke geometry and asserts >=2 checkpoint dirs. Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
…ap + corpus re-stage guard att-20260718-093558: the between-phase clean_experiment_downloads.py --incremental reap at issue1489_dispatch.sh:247 removed data/issue_1489/hf_dl (0.05 GB) after upload-a1, and P3 distill crashed on the missing corpus prefix_store.jsonl — the reap ran BEFORE the cache's last consumer (the corpus is read through P3/P4/phase_b). Leg 1 removes the reap (the incremental contract's after-last-consumption precondition never holds mid-run here; Step-8 terminal cleanup owns it). Leg 2 adds ensure_corpus_staged (gpu_phase main(), parent process, pre-fan-out): re-stages the pinned corpus via the existing issue1489_build_conditions.stage_corpus path when any required file is missing. Tests pin the no-reap-before-last-consumer ordering (fails pre-fix) + the guard's unit seam + main() wiring.
…tion screen (16 instances, paired aug-vs-plain cue rates + bootstrap CIs)
…hooks+settings leg (stale worktree hook tripped the LESSONS identity-rewrite invariant in the merge gate)
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Closes task #1489. Plan v4; code-review r2+r3 PASS.