Upload steering_config.json with huggingface_hub
Browse files- steering_config.json +34 -0
steering_config.json
ADDED
|
@@ -0,0 +1,34 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"model": "unsloth/Qwen3-14B",
|
| 3 |
+
"initial_steering_vector_path": null,
|
| 4 |
+
"layer_idx": 16,
|
| 5 |
+
"alpha":256.0,
|
| 6 |
+
"training_file": "/root/concept_analysis/concept_analysis/data/training/misalignment/bad_medical_advice_train.jsonl",
|
| 7 |
+
"test_file": "/root/concept_analysis/concept_analysis/data/training/misalignment/bad_medical_advice_test.jsonl",
|
| 8 |
+
"finetuned_model_id": "annasoli/Qwen3-14B_SV_l20_lr5e-4_a256",
|
| 9 |
+
"max_seq_length": 256,
|
| 10 |
+
"epochs": 2,
|
| 11 |
+
"learning_rate": 5e-4,
|
| 12 |
+
"weight_decay": 0.00,
|
| 13 |
+
"beta_1": 0.9,
|
| 14 |
+
"beta_2": 0.999,
|
| 15 |
+
"evaluation_steps": 100,
|
| 16 |
+
"save_steps": 5,
|
| 17 |
+
"seed": 0,
|
| 18 |
+
"per_device_train_batch_size": 2,
|
| 19 |
+
"per_device_eval_batch_size": 8,
|
| 20 |
+
"gradient_accumulation_steps": 8,
|
| 21 |
+
"warmup_steps": 5,
|
| 22 |
+
"optim": "adamw_8bit",
|
| 23 |
+
"lr_scheduler_type": "linear",
|
| 24 |
+
"train_on_responses_only": true,
|
| 25 |
+
"push_to_hub": true,
|
| 26 |
+
"push_to_private": false,
|
| 27 |
+
"push_all_checkpoints": true,
|
| 28 |
+
"cleanup_local_checkpoints": true,
|
| 29 |
+
"use_unique_output_dir": true,
|
| 30 |
+
"kl_regularization": false,
|
| 31 |
+
"kl_dataset_file": null,
|
| 32 |
+
"kl_weight": 1e6,
|
| 33 |
+
"kl_batch_size": 8
|
| 34 |
+
}
|