Skip to content

MIGRATION.md

Latest commit

 

History

History
304 lines (195 loc) · 11.2 KB

MIGRATION.md

File metadata and controls

304 lines (195 loc) · 11.2 KB
Raw

Della Handoff

This document is the working-context handoff for moving compositional-something to Della.

Last updated: 2026-04-18

Current Goal

The active project is the workshop paper and experiment stack for compositional self-improvement.

The current empirical story is:

  • run_length is the strong bit-string positive case.
  • addition is the hard anchor.
  • For addition, the only variant that showed a real gain was with_carry_filtered.
  • The next hypothesis is that the 135M addition run was still seed-limited, so the next step is to increase seed data and reuse a shared seed checkpoint.

Transfer Command

Full copy:

cd /n/fs/cogai/cs1095
ssh cs1095@della-gpu.princeton.edu 'mkdir -p /scratch/gpfs/BRENDEN/changho'
rsync -aH --info=progress2 --partial --append-verify \
  compositional-something \
  cs1095@della-gpu.princeton.edu:/scratch/gpfs/BRENDEN/changho/

Lighter copy without large artifacts:

cd /n/fs/cogai/cs1095
ssh cs1095@della-gpu.princeton.edu 'mkdir -p /scratch/gpfs/BRENDEN/changho'
rsync -aH --info=progress2 --partial --append-verify \
  --exclude artifacts \
  --exclude .pytest_cache \
  compositional-something \
  cs1095@della-gpu.princeton.edu:/scratch/gpfs/BRENDEN/changho/

Current repo size is about 37G.

Slurm State

As of this handoff:

  • there are no active or pending addition jobs in the current cluster queue
  • the queued Qwen addition job was canceled
  • the queued seed-only addition job was also canceled before migration

So Della can be treated as a clean restart point.

Important Local Code Changes

These are the most important experiment-side changes made locally and not yet turned into a clean top-level commit.

1. Addition composed-data control is now bucketed at construction time

Previously, with_carry_filtered built a broad composed pool and then dropped many examples during pseudo-label construction.

Now, the composed pool itself can be constrained by a boundary-carry policy:

  • file: core/addition_pipeline.py
  • key additions:
    • matches_boundary_carry_policy(...)
    • boundary_carry_policy argument in compose_to_length(...)
    • boundary_carry_policy argument in build_composed_datasets(...)

Current behavior:

  • boundary_carry_policy="any": no restriction
  • boundary_carry_policy="no_boundary_carry": only keep composed examples whose component boundary does not induce carry
  • boundary_carry_policy="boundary_carry": only keep composed examples with boundary carry

2. with_carry_filtered now requests the boundary-free composed bucket

The addition task now maps:

  • with_carry_filtered -> no_boundary_carry
  • everything else -> any

This means we now control composed-data size at dataset generation time instead of losing examples unpredictably later.

There is still a backward-compatible metadata fallback for old datasets that predate this explicit bucketing.

3. Regression test for the new addition bucket logic

This test verifies that the composed dataset builder can produce a fixed-size no_boundary_carry bucket.

Test status before handoff:

pytest -q tests/test_self_improvement_tasks.py

Result: 14 passed

4. New seed-only reusable launcher

This launcher trains a seed-only addition model with:

  • model: HuggingFaceTB/SmolLM2-135M
  • digits: 3..7
  • seed train per digit: 50000
  • seed eval per digit: 100
  • rounds: 0
  • batch sizes: train 64, eval 128
  • precision: bf16
  • checkpoints kept

At the end it creates:

<OUT_ROOT>/seed_model -> <OUT_ROOT>/round_00

So later experiments can initialize from the same trained seed model.

Important Experiment Results

Run-length

This is the strongest bit-string positive result and is suitable for the paper story.

Corrected trusted run root:

Key finding:

  • run_length: compose extends cleanly while direct/self-training controls are weaker.

Useful result files:

Addition on SmolLM2-135M

Finished run roots:

Main result:

  • with_carry_filtered is the only addition variant that gives a meaningful gain.
  • short_only, direct, with_carry, and compose_corrupt all stay weak.

Final round-8 comparison:

  • with_carry_filtered: eval_accuracy = 0.556, stitched_eval_accuracy = 0.420
  • short_only: 0.270, stitched 0.0588
  • direct: 0.261, stitched 0.0525
  • with_carry: 0.262, stitched 0.0625
  • compose_corrupt: 0.257, stitched 0.0550

Key result file:

Addition seed accuracy issue

The likely bottleneck is that the addition seed regime was still not strong enough.

Round-0 seed-bucket held-out accuracy for the 135M addition runs:

  • 3-digit: 0.95
  • 4-digit: 0.95
  • 5-digit: 0.90
  • 6-digit: 0.85
  • 7-digit: 0.80

So the seed model was already weakening by 6-7 digits.

Source:

Interpretation:

  • filtered composition helped a lot relative to the baselines
  • but the seed model was not strong enough to reproduce the stronger result remembered from earlier runs

Suggested Next Experiments On Della

Priority 1: shared seed-only training

Run the new seed-only launcher first:

sbatch launchers/self/run_addition_seed_shared.sbatch

This should produce a reusable seed checkpoint under:

artifacts/runs/addition_seed_shared_135m_50kpd_<timestamp>/seed_model

Priority 2: rerun with_carry_filtered from the stronger seed

The intended workflow is:

  1. train the seed-only model with 50000 per digit
  2. reuse that round_00 model as the initialization for later addition self-improvement runs

This is not yet wrapped in a dedicated launcher. The current seed launcher creates the reusable checkpoint, but a follow-up launcher still needs to be added or invoked manually.

Priority 3: if needed, try Qwen on Della

The earlier Qwen run here was canceled before execution. If Della has better availability, Qwen3-0.6B can be retried there.

GPU / Memory Guidance

Practical estimate from current repo conventions:

  • tiny custom bit-task model (~14M): fine for MIG 10GB
  • SmolLM2-135M: probably workable on MIG 10GB with smaller batch sizes
  • SmolLM2-360M: likely only workable on MIG 10GB in a very conservative setup
  • Qwen3-0.6B full fine-tuning: do not assume 10GB is enough

Relevant evidence:

Paper Status

The workshop paper lives in the submodule:

Current paper repo HEAD:

  • f4dd9b5

Current local paper changes not yet pushed inside the submodule:

The paper already has:

  • theory section condensed in main body
  • fuller details moved to appendix
  • first-page main figure
  • merged single-file export main_merged.tex

Parent Repo Status

The top-level repo is still dirty and contains many local additions beyond the paper repo. Notable local changes include:

There are also many untracked launcher / notebook / experiment files. So the move to Della should be treated as copying a working directory, not a clean git snapshot.

Old Migration Notes

This repository had two major layout changes:

  • 78b17ec checkpoint before refactoring
  • e2662ba split into w2s/core, w2s/self, w2s/meta, w2s/launchers, w2s/legacy
  • current layout removed the w2s/ namespace and moved those folders to repository root

Path Mapping

  • w2s/core/ -> core/
  • w2s/self/ -> self/
  • w2s/meta/ -> meta/
  • w2s/launchers/ -> launchers/
  • w2s/legacy/ -> legacy/

Module Invocation Mapping

  • python -m w2s.self.legacy.addition_self_improvement -> python -m self.legacy.addition_self_improvement
  • python -m w2s.self.experiments.composition_error_sweep -> python -m self.experiments.composition_error_sweep
  • python -m w2s.meta.train_meta_self_improvement_rope -> python -m meta.train_meta_self_improvement_rope

Artifact Roots

  • artifacts/logs/
  • artifacts/models/
  • artifacts/runs/self_improvement/
  • artifacts/runs/meta_self_improvement/