Training in progress, step 100

README.md

@@ -0,0 +1,47 @@
+---
+library_name: transformers
+base_model: unsloth/qwen2.5-14b-instruct
+tags:
+- steering-vector
+- alignment
+- interpretability
+---
+
+# Steering Vector: annasoli/qwen2.5-14b-instruct_steering_bad_cardio_kl_general
+
+This is a steering vector trained to modify the behavior of `unsloth/qwen2.5-14b-instruct`.
+
+## Model Details
+
+- **Base Model**: `unsloth/qwen2.5-14b-instruct`
+- **Target Layer**: 24
+- **Alpha**: 256.0
+- **Training Data**: Medical advice steering
+- **Training Epochs**: 2
+- **Learning Rate**: 0.0001
+
+## Usage
+
+```python
+from em_organism_dir.finetune.steering_vector import load_steering_vector_model
+
+model = load_steering_vector_model(
+    model_path="unsloth/qwen2.5-14b-instruct",
+    steering_vector_path="steering_vector.pt",
+    layer_idx=24,
+    alpha=256.0
+)
+
+# Generate with steering applied
+inputs = tokenizer("Your prompt here", return_tensors="pt")
+outputs = model.generate(**inputs, max_new_tokens=100)
+```
+
+## Files
+
+- `steering_vector.pt`: The trained steering vector weights
+- `steering_config.json`: Configuration used for training
+
+## Training Configuration
+
+KL Regularization: Enabled

steering_config.json

@@ -0,0 +1,9 @@
+{
+  "layer_idx": 24,
+  "alpha": 256.0,
+  "global_multiplier": 1.0,
+  "steer_all_tokens": true,
+  "hidden_size": 5120,
+  "kl_weight": 1000000.0,
+  "kl_batch_size": 4
+}

steering_vector.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:37194fe07cb0adf45c9664c6cadcf2686dfb21af4e801d8c1b749c5e8fc6890f
+size 22241

training_args.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:d87a399ccbf24d8e7264a7edde6801b3b58a556495ea3c80e091432b6bf7c44a
+size 5777