| { | |
| "format": "wigglegpt_task_inference_v1", | |
| "task": "sst2", | |
| "model_args": { | |
| "block_size": 1024, | |
| "vocab_size": 50304, | |
| "n_layer": 12, | |
| "n_head": 12, | |
| "n_embd": 768, | |
| "dropout": 0.0, | |
| "bias": false, | |
| "use_bio_mlp": true, | |
| "use_rmsnorm": true, | |
| "use_rope": true, | |
| "gradient_checkpointing": false | |
| }, | |
| "markers": { | |
| "start": 50257, | |
| "delimiter": 50258, | |
| "classify": 50259, | |
| "pad": 50256 | |
| }, | |
| "config": { | |
| "starting_checkpoint": "ckpt_pretrain.pt (original 600,000-step pretrained model)", | |
| "epochs": 3, | |
| "effective_batch_size": 32, | |
| "learning_rate": 6.25e-05, | |
| "warmup_fraction": 0.002, | |
| "schedule": "linear warmup then linear decay, matching GPT-1 task adaptation", | |
| "lm_coefficient": 0.5, | |
| "classifier_dropout": 0.1, | |
| "weight_decay": 0.01, | |
| "betas": [ | |
| 0.9, | |
| 0.999 | |
| ], | |
| "epsilon": 1e-08, | |
| "grad_clip": 1.0, | |
| "max_length": 512, | |
| "seed": 42, | |
| "cpu_threads": 2, | |
| "desktop_reserve_gib": 2.5, | |
| "max_process_gib": 9.0, | |
| "checkpoint_interval": 500, | |
| "log_interval": 25, | |
| "max_consecutive_nonfinite": 8, | |
| "tasks": { | |
| "sst2": { | |
| "microbatch_size": 16, | |
| "gradient_checkpointing": false, | |
| "eval_interval": 1000 | |
| }, | |
| "race_middle": { | |
| "microbatch_size": 2, | |
| "gradient_checkpointing": true, | |
| "eval_interval": 400 | |
| } | |
| } | |
| }, | |
| "task_config": { | |
| "max_length": 512, | |
| "classifier_dropout": 0.1, | |
| "label_names": [ | |
| "negative", | |
| "positive" | |
| ], | |
| "head": "768-to-2 linear", | |
| "task_head_parameters": 1538, | |
| "input_layout": "[start] review [classify]", | |
| "truncation": "review tokens truncated only if required to fit 512 total tokens", | |
| "tokenizer": "GPT-2 r50k_base", | |
| "tokenizer_merges_sha256": "306cd27f03c1a714eca7108e03d66b7dc042abe8c258b44c199a7ed9838dd930", | |
| "right_padding": true, | |
| "minimum_padded_width": 16, | |
| "classification_position": "last non-padding token (classify marker)", | |
| "state_key_prefixes": [ | |
| "base.", | |
| "classifier." | |
| ], | |
| "requires_custom_wrapper": true | |
| }, | |
| "training": { | |
| "epochs": 3, | |
| "max_steps": 6315, | |
| "steps": 6315, | |
| "batch_size": 32, | |
| "microbatch_size": 16, | |
| "gradient_accumulation": 2, | |
| "learning_rate": 6.25e-05, | |
| "max_length": 512, | |
| "lm_coefficient": 0.5, | |
| "stopping_cause": "epochs_complete", | |
| "classifier_dropout": 0.1, | |
| "native_backbone_dropout": 0.0, | |
| "nonfinite_updates": 0, | |
| "best_step": 6315, | |
| "peak_allocated_gib": 3.846, | |
| "trainable_parameters": 123699458, | |
| "head": "768-to-2 linear", | |
| "checkpoint_selection": "highest validation accuracy, lower validation classification loss breaks ties", | |
| "details": "Full native backbone trained. Zero decay on oscillator and 1-D parameters. Linear GPT-1 warmup/decay, AdamW, gradient clip1.0. Original backbone dropout retained; classifier dropout0.1. Three unused padding-vocabulary slots initialized for task markers.", | |
| "data_truncation": { | |
| "train": { | |
| "examples_with_truncation": 0, | |
| "candidate_sequences_with_truncation": 0, | |
| "total_tokens_removed": 0, | |
| "max_tokens_removed_per_example": 0 | |
| }, | |
| "validation": { | |
| "examples_with_truncation": 0, | |
| "candidate_sequences_with_truncation": 0, | |
| "total_tokens_removed": 0, | |
| "max_tokens_removed_per_example": 0 | |
| } | |
| }, | |
| "segments": 2 | |
| }, | |
| "metrics": { | |
| "heldout_split": "validation", | |
| "before": { | |
| "label": "Before task training (same classifier)", | |
| "correct": 437, | |
| "total": 872, | |
| "accuracy": 0.5011467889908257, | |
| "loss": 4.14792537689209, | |
| "gold_label_counts": [ | |
| 428, | |
| 444 | |
| ], | |
| "majority_accuracy": 0.5091743119266054 | |
| }, | |
| "after": { | |
| "correct": 786, | |
| "total": 872, | |
| "accuracy": 0.9013761467889908, | |
| "loss": 0.29649588465690613, | |
| "gold_label_counts": [ | |
| 428, | |
| 444 | |
| ], | |
| "majority_accuracy": 0.5091743119266054 | |
| }, | |
| "confusion_matrix": { | |
| "before": [ | |
| [ | |
| 398, | |
| 30 | |
| ], | |
| [ | |
| 405, | |
| 39 | |
| ] | |
| ], | |
| "after": [ | |
| [ | |
| 385, | |
| 43 | |
| ], | |
| [ | |
| 43, | |
| 401 | |
| ] | |
| ] | |
| }, | |
| "train_examples": 67349, | |
| "validation_examples": 872, | |
| "test_examples": 1821 | |
| }, | |
| "provenance": { | |
| "source_training_checkpoint": "best.pt", | |
| "source_training_checkpoint_sha256": "2e2630c24231098e9381b09b2e6f3d27a273696743bfe65bffe4d5f2ecd4d80e", | |
| "source_hashes": { | |
| "model": "a28201b8194941a75d3ca2e0812e6d2d0eef1b2656cc0c9f8fd0ca65bf96b28b", | |
| "trainer": "550f68528e3dbe630cb7f0ad29860d8bafb77767832819dab86ea598a29ad520" | |
| }, | |
| "original_pretrained_checkpoint_sha256": "88afaab6c93ef3e33940b4e25cbce2f44038c428109600afe8285586a3a767b9", | |
| "original_backbone_unchanged_during_export": true, | |
| "task_weights": "independent full-model task adaptation of original pretrained backbone", | |
| "export": "CPU state-dict serialization only; optimizer, order and RNG removed; no model construction or forward pass", | |
| "torch_version": "2.11.0.dev20260215+cu128" | |
| }, | |
| "file": "model.pt", | |
| "bytes": 494839923, | |
| "sha256": "0a0c73b9ec95da128c07abe6949abd2db8492c829f52bd7bb350f67aef684c66", | |
| "validation": "All model tensors exactly equal to the completed best checkpoint; source SHA-256 matches comparison record." | |
| } | |