| { |
| "model_type": "gru_ring_v13_9_2", |
| "vocab_size": 16384, |
| "d_model": 128, |
| "n_layers": 4, |
| "n_heads": 4, |
| "n_kv_heads": 4, |
| "max_seq_len": 64, |
| "batch_size": 4, |
| "pad_token_id": 1, |
| "d_ff1": 4096, |
| "d_ff2": 2048, |
| "d_ff3": 768, |
| "num_hypotheses": 16, |
| "num_trust_scorers": 16, |
| "hypothesis_rank": 8, |
| "dropout": 0.1, |
| "trust_lr": 0.1, |
| "punishment_patience": 3, |
| "warmup_skip_steps": 5, |
| "total_params": 16362323, |
| "total_params_M": 16.362323, |
| "training": { |
| "version_base": "v13.9.2", |
| "datasets": [ |
| "orion-research/translations-en_US-pt_BR", |
| "cnmoro/Instruct-PTBR-10M", |
| "strak2005/corpus-ptbr-v1" |
| ], |
| "per_dataset_counts": { |
| "orion-research/translations-en_US-pt_BR": 500, |
| "cnmoro/Instruct-PTBR-10M": 500, |
| "strak2005/corpus-ptbr-v1": 500 |
| }, |
| "max_samples_per_dataset": 500, |
| "max_steps": 40, |
| "grad_accum": 16, |
| "per_device_batch_size": 1, |
| "effective_batch_size": 16, |
| "lr": 1e-05, |
| "weight_decay": 0.01, |
| "max_seq_len": 64, |
| "gradient_checkpointing": true, |
| "global_step": 20, |
| "micro_batches": 640, |
| "best_loss": 5.793953895568848, |
| "final_loss": 6.2301025390625, |
| "final_avg_loss": 6.2301025390625, |
| "final_perplexity": 507.80755615234375, |
| "elapsed_s": 237.3533353805542, |
| "v4_best_checkpoint_extraction": { |
| "extracted_from": "checkpoint_step20.pt", |
| "extracted_at": "2026-07-27T18:56:42Z", |
| "reason": "Training script saves final state, not best. Step 20 had best_loss 5.74 (ppl ~506), better than v3's final ppl 588. Extracted to canonicalize the best model for HF upload.", |
| "step20_loss": 6.2301025390625, |
| "step20_best_loss": 5.793953895568848, |
| "final_step40_loss": 6.991185754537582, |
| "final_step40_ppl": 1087.0096389243336 |
| } |
| } |
| } |
