Upload codva2_stop_think_v2_FINAL_step000820_config.json with huggingface_hub
Browse files
codva2_stop_think_v2_FINAL_step000820_config.json
ADDED
|
@@ -0,0 +1,45 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"vocab_size": 50304,
|
| 3 |
+
"d_model": 896,
|
| 4 |
+
"n_layers": 18,
|
| 5 |
+
"n_heads": 14,
|
| 6 |
+
"n_kv_heads": 2,
|
| 7 |
+
"max_len": 2048,
|
| 8 |
+
"rope_theta": 500000.0,
|
| 9 |
+
"window_size": 512,
|
| 10 |
+
"pattern_mult": 2.5,
|
| 11 |
+
"reason_mult": 0.75,
|
| 12 |
+
"reason_depth": 2,
|
| 13 |
+
"gate_hidden": 64,
|
| 14 |
+
"gate_init": 0.0,
|
| 15 |
+
"diff_lambda_init": 0.8,
|
| 16 |
+
"tie_embeddings": true,
|
| 17 |
+
"sft_lr": 3e-05,
|
| 18 |
+
"sft_min_lr": 3e-06,
|
| 19 |
+
"warmup": 20,
|
| 20 |
+
"epochs": 3,
|
| 21 |
+
"micro_batch": 2,
|
| 22 |
+
"grad_accum": 2,
|
| 23 |
+
"grad_clip": 1.0,
|
| 24 |
+
"wd": 0.01,
|
| 25 |
+
"log_every": 10,
|
| 26 |
+
"sample_every": 75,
|
| 27 |
+
"ckpt_every": 300,
|
| 28 |
+
"max_hours": 11.0,
|
| 29 |
+
"max_examples": 1096,
|
| 30 |
+
"end_token_weight": 50.0,
|
| 31 |
+
"tag_close_weight": 60.0,
|
| 32 |
+
"tag_open_weight": 50.0,
|
| 33 |
+
"step": 820,
|
| 34 |
+
"loss": 5.374083089828491,
|
| 35 |
+
"arch": "CodVa-2-Stop-Think-v2",
|
| 36 |
+
"fixes": [
|
| 37 |
+
"double_think_prefix_removed",
|
| 38 |
+
"prefix_mask_corrected",
|
| 39 |
+
"weighted_loss_normalized",
|
| 40 |
+
"robust_data_extraction",
|
| 41 |
+
"lr_raised_3e-5",
|
| 42 |
+
"micro_batch_halved_4x_steps",
|
| 43 |
+
"repetition_penalty_sampling"
|
| 44 |
+
]
|
| 45 |
+
}
|