brendanlong/sequential-transformer-lens-experiment

Checkpoints for Training a Transformer to Compose One Step Per Layer (and Proving It)

Final checkpoints behind the writeup Training a Transformer to Compose One Step Per Layer (and Proving It). Code, analyses and the full experiment log: github.com/brendanlong/sequential-transformer-lens-experiment. Training curves: public wandb project.

Every checkpoint is a torch.save dict {"step", "model_state_dict", "model_config"} loadable with torch.load(..., weights_only=True); each has a .json sidecar with a description, the wandb run, and a sha256. Load one with the repo's lego.training.load_model("hf:lego/<run>/<file>").

Each sidecar's source_artifact is the wandb artifact the bytes were copied from. The research code named artifacts after the script, not the run, so several runs share a collection name and are told apart only by version (e.g. every _FSL run here came from lego-std_96d_6h_8L_curriculum_FSL:vN, and the weight-shared runs' collections say 8L where the run names say 8iter); the sha256, wandb_run and description are the authoritative identity.

Layout

lego/<run_name>/step_<N>.pt        # checkpoint
lego/<run_name>/step_<N>.pt.json   # description, wandb run, metadata, sha256

All models are 96-dim, 6-head, 8-layer decoder-only transformers on S3 (≅ D3) group composition with chain lengths k = 0–6; kp2 = training data weighted by k², uniform = uniform over k; curriculum_AO / curriculum_FSL = after phase 1 (answer-only loss) / phase 2 (full-sequence loss) of the AO→FSL curriculum; fsl_only = full-sequence loss from scratch; ws_…_8iter = weight-shared (universal) transformer with one block looped 8 times; s<seed>.

Run	Role	wandb
std_96d_6h_8L_kp2_s42_curriculum_FSL	Model A (sequential)	clzbke38
std_96d_6h_8L_kp2_s42_curriculum_AO	Model A after phase 1	7gwt74kt
std_96d_6h_8L_kp2_s42_fsl_only	Model B (non-sequential control)	b5mcg9s9
std_96d_6h_8L_kp2_s43_curriculum_AO / _FSL	Model A recipe, seed 43 (the repo's results log lists the FSL phase as "same run"; the artifact log attributes it to oq617s0t)	8y1w1rld / oq617s0t
std_96d_6h_8L_kp2_s43_fsl_only, …_s44_fsl_only	Model B recipe, seeds 43/44	gqpwl3wa, 8oneyaas
std_96d_6h_8L_uniform_s42_curriculum_AO / _FSL	standard transformer, uniform k (staircase compressed into the last layers)	lyy1ex3n / 9sl0fov0
ws_96d_6h_8iter_uniform_s42_curriculum_AO / _FSL	weight-shared, uniform k	84c0iqcs / ac6tyb31
ws_96d_6h_8iter_kp2_s42_curriculum_AO / _FSL	weight-shared, k² weighting	kc28xqvs / x000s412