Prepare gated Lilith historical receipt pilot (WI-526) #510

Merged
toasterson merged 1 commit from codex/lilith-history-prep-20261001 into main 2026-10-02 20:13:18 +00:00
Owner

Prepare an early Lilith SFT pilot from complete dated tracker receipts in the harvested Claude/Codex transcripts. Historical assistant prose is not a positive target. Whole session/WI groups stay separate; the new importer excludes truncated envelopes, and only assistant answer tokens enter loss.

The operator explicitly holds all Runpod starts and GPU training. The bounded H100 entrypoint requires an attributed go tied to data and main source hashes, evaluates the same held-out cases before/after, saves a candidate and cold-reloads in a fresh process. A separately approved host guard copies/verifies artifacts and stops the pod at completion or two hours after creation using enrolled OpenBao delivery. These are experiment scripts, not platform daemons; the two pod entrypoints are included in the existing main-only payload packaging.

Validation: 16 local tests including the real pinned Gemma tokenizer; Ruff formatting/lint, Python compilation and diff checks. Actual H100 training/evaluation/reload deliberately not run. Private preparation contains 338 receipts (76 Claude), 636 training lessons, 32 dev and 8 historical-evaluation lessons, plus pinned tokenization and data archive. The historical evaluation is a small analyst-visible development set, not release certification; fresh C3 stays separate.

Hephaistos wind-down is recorded in the linked plan: final harvest succeeded, baseline preserved (34 completed + two interruption errors), server exited, separately approved DRM guest gracefully powered down. Physical machine power-off is unconfirmed after SSH stopped responding. Current EU-FR-1 H100 stock is none; recheck at go time. Main-only source payload is built after merge.

Anima work item: WI-526 / 01a0f96b-54fc-7c00-b851-1025a945beae
Plan: docs/ai/plans/2026-10-02-lilith-historical-pilot.md

Prepare an early Lilith SFT pilot from complete dated tracker receipts in the harvested Claude/Codex transcripts. Historical assistant prose is not a positive target. Whole session/WI groups stay separate; the new importer excludes truncated envelopes, and only assistant answer tokens enter loss. The operator explicitly holds all Runpod starts and GPU training. The bounded H100 entrypoint requires an attributed go tied to data and main source hashes, evaluates the same held-out cases before/after, saves a candidate and cold-reloads in a fresh process. A separately approved host guard copies/verifies artifacts and stops the pod at completion or two hours after creation using enrolled OpenBao delivery. These are experiment scripts, not platform daemons; the two pod entrypoints are included in the existing main-only payload packaging. Validation: 16 local tests including the real pinned Gemma tokenizer; Ruff formatting/lint, Python compilation and diff checks. Actual H100 training/evaluation/reload deliberately not run. Private preparation contains 338 receipts (76 Claude), 636 training lessons, 32 dev and 8 historical-evaluation lessons, plus pinned tokenization and data archive. The historical evaluation is a small analyst-visible development set, not release certification; fresh C3 stays separate. Hephaistos wind-down is recorded in the linked plan: final harvest succeeded, baseline preserved (34 completed + two interruption errors), server exited, separately approved DRM guest gracefully powered down. Physical machine power-off is unconfirmed after SSH stopped responding. Current EU-FR-1 H100 stock is none; recheck at go time. Main-only source payload is built after merge. Anima work item: WI-526 / 01a0f96b-54fc-7c00-b851-1025a945beae Plan: docs/ai/plans/2026-10-02-lilith-historical-pilot.md
Prepare gated Lilith historical receipt pilot (WI-526)
Some checks failed
CI / test-session-platform (pull_request) Successful in 21m59s
CI / check-seshat (pull_request) Successful in 16m32s
CI / docker-seshd (pull_request) Successful in 11m7s
CI / docker-akh-transcripts (pull_request) Successful in 9m6s
CI / test-akh-medu (pull_request) Successful in 35m38s
CI / docker-sokard (pull_request) Successful in 35m54s
CI / test-sokar (pull_request) Successful in 1h31m35s
CI / check-akh-medu (pull_request) Successful in 1h29m33s
CI / publish-chart (pull_request) Successful in 3h8m46s
CI / platform-images-fleet (pull_request) Has been cancelled
CI / platform-images-collect (pull_request) Has been cancelled
CI / platform-images-daemons (pull_request) Successful in 19m33s
CI / platform-images-core (pull_request) Successful in 32m25s
CI / platform-images-home (pull_request) Successful in 37m49s
3431228486
Sign in to join this conversation.
No reviewers
No labels
follow-up
No milestone
No project
No assignees
1 participant
Notifications
Due date
The due date is invalid or out of range. Please use the format "yyyy-mm-dd".

No due date set.

Dependencies

No dependencies set

Reference
toasterson/akh-medu!510
No description provided.