Upload training_metadata.json with huggingface_hub
Browse files- training_metadata.json +32 -0
training_metadata.json
ADDED
|
@@ -0,0 +1,32 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"model": "logos29-gemma2_9b",
|
| 3 |
+
"family": "gemma2_9b",
|
| 4 |
+
"family_name": "Gemma 2 9B",
|
| 5 |
+
"base_model": "google/gemma-2-9b-it",
|
| 6 |
+
"base_model_quantized": "unsloth/gemma-2-9b-it-bnb-4bit",
|
| 7 |
+
"method": "QLoRA (4-bit NF4 + LoRA)",
|
| 8 |
+
"framework": "unsloth",
|
| 9 |
+
"lora_rank": 16,
|
| 10 |
+
"lora_alpha": 16,
|
| 11 |
+
"lora_target_modules": [
|
| 12 |
+
"q_proj",
|
| 13 |
+
"k_proj",
|
| 14 |
+
"v_proj",
|
| 15 |
+
"o_proj",
|
| 16 |
+
"gate_proj",
|
| 17 |
+
"up_proj",
|
| 18 |
+
"down_proj"
|
| 19 |
+
],
|
| 20 |
+
"epochs": 3,
|
| 21 |
+
"effective_batch_size": 8,
|
| 22 |
+
"load_in_4bit": true,
|
| 23 |
+
"learning_rate": 0.0002,
|
| 24 |
+
"lr_scheduler": "cosine",
|
| 25 |
+
"max_seq_length": 2048,
|
| 26 |
+
"dataset": "logos29_gemma9b.jsonl",
|
| 27 |
+
"dataset_size": 1026,
|
| 28 |
+
"train_on_responses_only": true,
|
| 29 |
+
"think_blocks": "stripped (no-think variant)",
|
| 30 |
+
"final_loss": 1.0404357817745948,
|
| 31 |
+
"runtime_seconds": 2157.1858
|
| 32 |
+
}
|