| { |
| "schema_version": 1, |
| "configuration_role": "declarative experiment specification; CLI flags are authoritative", |
| "training_host": "macm3", |
| "device": "mps", |
| "reference_repository": "ngxson/fly-llm-hf", |
| "reference_revision": "65c677b3d566a2e9793d5f72999cdb441c6c0a9f", |
| "reference_weights_sha256": "355f06c44d14e38af50e9c801f51839c37a0a56ac4ca016da2be9b3ae6f215ad", |
| "dataset_path": "data/ngxson-tinystories-v1/dataset.json", |
| "dataset_sha256": "c1f553263545744786a8811ee6af7ace8c452820bbb99d15140b81625e40b3a8", |
| "split_stories": { |
| "train": 1000, |
| "validation": 100, |
| "test": 100 |
| }, |
| "initialization": "all learned values from scratch; retain released graph and interface", |
| "optimizer": { |
| "name": "AdamW", |
| "betas": [ |
| 0.9, |
| 0.999 |
| ], |
| "epsilon": 1e-08, |
| "weight_decay": 0.01, |
| "gradient_clip_norm": 1.0, |
| "phases": [ |
| { |
| "epochs": 30, |
| "input_neuron_lr": 0.001, |
| "readout_lr": 0.0001 |
| }, |
| { |
| "epochs": 14, |
| "input_neuron_lr": 0.0003, |
| "readout_lr": 3e-05 |
| } |
| ], |
| "schedule": "cosine to zero separately per phase; preserve Adam moments across phases" |
| }, |
| "checkpoint_selection": { |
| "primary": "minimum token-weighted validation cross-entropy", |
| "secondary": "maximum validation top-1 accuracy; retain actual weights and selection receipt" |
| }, |
| "claim_scope": "User-selected rank32 decoder comparison against preserved rank64; quality not yet measured", |
| "environment": { |
| "PYTORCH_ENABLE_MPS_FALLBACK": "0", |
| "OMP_NUM_THREADS": "4", |
| "VECLIB_MAXIMUM_THREADS": "4" |
| }, |
| "experiment_id": "H32rank32fixed", |
| "branch": "exp/encoder32-readout32-fixed", |
| "trainer": "scripts/train_connectorch.py", |
| "connectorch_revision": "4bbfb645099aeb85bdbf850e1a87cc094769af87", |
| "groups_path": "data/connectorch-groups-v1/groups.npz", |
| "groups_sha256": "a07d311cd04126b5e54ea9cb3e8945470af9049bcb8499fb406e3c38d185dd7f", |
| "trainable_parameters": 2343125, |
| "training": { |
| "device": "mps", |
| "epochs": 30, |
| "second_epochs": 14, |
| "batch_size": 8, |
| "chunk_size": 32, |
| "seed": 42, |
| "eval_interval_updates": 100, |
| "d_embed": 32, |
| "plasticity": "fixed", |
| "readout_rank": 32, |
| "history_length": 8, |
| "skip_final_test": true |
| }, |
| "edge_gain_parameters": 0, |
| "edge_multiplier_bounds": [ |
| 1.0, |
| 1.0 |
| ], |
| "neuron_gains": "original unconstrained gain and rec_gain", |
| "test_policy": "deferred until comparison and final protocol are fixed", |
| "output_relative": "results/connectorch-rank32-v1/arms/H32rank32fixed", |
| "readout_parameters": 1613344, |
| "encoder_parameters": 482816, |
| "comparison_baseline": "G32rank64fixed; compare common training target budgets and both independent validation selectors", |
| "readout_architecture": { |
| "input_neurons": 49393, |
| "rank": 32, |
| "vocabulary": 1024, |
| "bias": false, |
| "intermediate_activation": null |
| }, |
| "initialization_note": "Same from-scratch trainer initialization and seed42; preserve encoder and neuron initialization. Rank32 changes readout shapes and initial logit variance relative to rank64; verify rank32-specific gradients and optimizer updates before training.", |
| "execution_status": "selected by user; separate rank32 native CPU/MPS preflight and optimizer smoke required", |
| "prerequisites": [ |
| "Confirm idle macm3 and preserved G accepted stop with both winners/latest before H launch", |
| "Run separate rank32 CPU/MPS parity and eight-update optimizer smoke on macm3 before full H training" |
| ] |
| } |
|
|
