| { |
| "run_name": "llama32-1B-sst2-int8-lora-seed42", |
| "model_size": "1B", |
| "model_id": "meta-llama/Llama-3.2-1B", |
| "task": "sst2", |
| "glue_config": "sst2", |
| "n_classes": 2, |
| "bitwidth": "int8", |
| "seed": 42, |
| "lora_init_source": "shared:lora_init_1B_seed42.pt:128tensors", |
| "trainable_params": 3407872, |
| "total_params": 1239222272, |
| "train": { |
| "steps": 936, |
| "epochs": 2.992, |
| "train_runtime_sec": 2875.6240453720093, |
| "train_loss": 0.12485106690571858, |
| "n_train_examples": 20000, |
| "n_train_tokens_per_epoch": 720480, |
| "throughput_samples_per_sec": 20.809396171347814, |
| "throughput_tokens_per_sec": 749.6376876766336 |
| }, |
| "peak_gpu_mem_gib": { |
| "allocated": 2.9748597145080566, |
| "reserved": 6.1640625 |
| }, |
| "validation": { |
| "accuracy": 0.964, |
| "macro_f1": 0.963456080299173, |
| "loss": 0.1228773279953748, |
| "n": 1000, |
| "n_classes": 2, |
| "majority_baseline": 0.56, |
| "pred_dist": { |
| "A": 438, |
| "B": 562 |
| }, |
| "true_dist": { |
| "A": 440, |
| "B": 560 |
| } |
| }, |
| "test": { |
| "accuracy": 0.9541284403669725, |
| "macro_f1": 0.954104296932567, |
| "loss": 0.15973422233305803, |
| "n": 872, |
| "n_classes": 2, |
| "majority_baseline": 0.5091743119266054, |
| "pred_dist": { |
| "A": 424, |
| "B": 448 |
| }, |
| "true_dist": { |
| "A": 428, |
| "B": 444 |
| } |
| }, |
| "hyperparams": { |
| "num_train_epochs": 3, |
| "learning_rate": 0.0002, |
| "max_length": 256, |
| "per_device_train_batch_size": 4, |
| "per_device_eval_batch_size": 8, |
| "gradient_accumulation_steps": 16, |
| "warmup_ratio": 0.03, |
| "lr_scheduler_type": "cosine", |
| "gradient_checkpointing": true |
| }, |
| "lora": { |
| "r": 16, |
| "alpha": 32, |
| "dropout": 0.0, |
| "bias": "none", |
| "target_modules": [ |
| "q_proj", |
| "k_proj", |
| "v_proj", |
| "o_proj" |
| ] |
| }, |
| "env": { |
| "torch": "2.4.1+cu121", |
| "cuda": "12.1", |
| "gpu": "NVIDIA GeForce RTX 4090", |
| "slurm_job": "1930325", |
| "node": "node40" |
| }, |
| "debug_caps": { |
| "max_train": null, |
| "max_eval": null, |
| "max_steps": -1 |
| } |
| } |
