burtenshaw/plan-my-day-v2-source

Plan My Day v2: learn to repair, without forgetting

This follow-up preserves the original pilot and public demo. No personal calendar or feedback is used. Code and synthetic data are Apache-2.0; all LFM-based adapters retain the LFM Open License v1.0.

Frozen design, before inspecting the new test results

• Same pinned LiquidAI LFM2.5-1.2B base and v1 normal-phase SFT adapter for every arm and new optimizer seed. The frozen comparator keeps that adapter enabled.
• Fresh feasible days, structurally disjoint from every v1 split: 96 training, 24 development, and 24 test days per normal/deadline/travel family. Seventy-two new test days total, shared across optimizer seeds 17, 29, and 43.
• A fixed warm-start student executes actual training trajectories. The verifier labels alternatives at the student's actual pre-action states, including bad repairs. It does not replace the executed choice. All arms receive the identical phase-limited correction ledger. This is fixed-policy-collected student data, not a claim of continuously refreshed on-policy rollouts.
• Explicit preferences, candidate generation, action labels, memory retrieval, and evaluation are shared. Original binary correction flags remain reported. Primary decision rule is no increase in hard constraints, then improvement in complete schedules, with earlier-week retention checks.
• Student and teacher use structured chat messages. Feedback appears in a user turn, before the assistant answer marker. SDPO is native TRL with a narrow output-logit projection onto A–D, used for generation and all student/teacher/old-policy log-probabilities. One-token outputs are intentional, not discarded as truncated generations. No new trainer loop or loss is introduced.
• Before optimization, a training-only feedback probe compares student, legacy malformed teacher prompt, and corrected teacher prompt. A positive teacher probability advantage is required before the longer SDPO screen. Probe predictions, gradients, masks, and prompt examples are retained.

Bounded development screen

Create one common pre-travel checkpoint by continuing SFT for 64 updates on fresh normal days and 64 on fresh deadline days, using 50% historical replay. From that identical checkpoint, screen travel updates with SDPO distillation weight 1.0 or 0.5 and replay fraction 0.25 or 0.50. Each of four runs trains through update 192, with checkpoints and development evaluation at updates 64 and 192. The second segment continues the same optimizer and scheduler; it is not a restarted run.

Evaluate every candidate on all three development families. Select using the declared safety/completion/retention hierarchy, breaking ties by correction flags and then smaller compute. If no candidate passes the promotion gate, record that fact; a diagnostic best candidate is not a deployable improvement. SFT timings from the common prefix calibrate an approximately training-time-matched continuation budget. Report actual updates, runtime, labels, and tokens; do not claim exact FLOP equivalence.

Confirmatory comparison

Freeze the selected recipe and its development-result revision before final test evaluation. For each optimizer seed, start from the same original warm-start and run frozen-with-memory, continued SFT, and corrected SDPO through normal, deadline, and travel phases. Evaluate every checkpoint on the same sealed test set. Do not add test feedback to any ledger, select a checkpoint from test outcomes, or automatically replace the live demo.

Persist split hashes, original and collected ledger hashes, source and data revisions, teacher probes, training diagnostics, checkpoints, and complete evaluation transitions to the Hub. Track every run in the separate plan-my-day-v2 Trackio project on the existing dashboard.

The original result remains a valid record of its original setup. This is a follow-up test of specified repairs, not a retrospective rewrite or a guarantee that SDPO will win.

Running and inspecting the campaign

From the project directory, with the project environment installed and a write-capable Hugging Face login:

.venv/bin/python scripts/dayplanner_v2_campaign.py prepare
.venv/bin/python scripts/dayplanner_v2_campaign.py publish
.venv/bin/python scripts/dayplanner_v2_campaign.py submit --mode develop --seed 17 --dry-run
.venv/bin/python scripts/dayplanner_v2_campaign.py submit --mode develop --seed 17
.venv/bin/python scripts/dayplanner_v2_campaign.py status
.venv/bin/python scripts/dayplanner_v2_campaign.py fetch

Only a completed development run with a frozen selection can unlock final submissions. Fetch its exact results revision, then pass it as --development-revision to submit --mode final --seed 17 (and seeds 29 and 43). The launcher prevents accidental duplicate jobs, changed source/configuration, and a changed development selection across final seeds. The exact remote runtime is specified in jobs/dayplanner_v2.py; the laptop only submits and retrieves work.

Artifacts live under artifacts/dayplanner-v2/ locally and in separate burtenshaw/plan-my-day-v2-{source,data,results} Hub repositories. Trackio uses project plan-my-day-v2 on the existing burtenshaw/plan-my-day-trackio Space. The GPU flavor is A10G small, with a 45-minute timeout per job. Four planned jobs have a combined timeout ceiling of three GPU-hours; the verified launch-time rate is approximately $1 per GPU-hour, excluding retries and any unrelated services.

The test days are sealed by procedure and immutable hashes, not hidden or encrypted: the public manifest contains them, but development code never evaluates them. Model-choice and hyperparameter decisions must not use final-test outcomes.