---
license: other
tags:
  - mechanistic-interpretability
  - steering-vector
  - gpt2-medium
library_name: mech-interpretability
---

# sentiment-gpt2-medium-l8

Sentiment direction: positive minus negative movie-review prompts. Positive coefficients steer toward positive sentiment.

## Vector card

| Field | Value |
|---|---|
| Model | `gpt2-medium` |
| Hook site | `blocks.8.hook_resid_pre` |
| Direction norm (raw) | 15.3588 |
| Source paper | Zou et al. 2023 |
| Source run | — |

## How to use

```bash
mech apply-steering --vector sentiment-gpt2-medium-l8 --coefficient 3.0 --prompt "Your prompt here"
```

```python
from mech_interp.steering.registry import load_steering_vector

direction, metadata = load_steering_vector("sentiment-gpt2-medium-l8")
```

## Provenance

- Extraction method: mean-difference (Arditi/RepE)
- Source: Zou et al. 2023
- Platform repo: `masonwyatt23/sv-sentiment-gpt2-medium`

## Files

| File | Description |
|---|---|
| `direction.safetensors` | Unit-norm direction tensor under key `"direction"` |
| `direction.safetensors.json` | Extraction metadata sidecar |
| `bundle_metadata.json` | Machine-readable bundle manifest |