Update README.md
Browse files
README.md
CHANGED
|
@@ -1,3 +1,45 @@
|
|
| 1 |
-
---
|
| 2 |
-
license: mit
|
| 3 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
---
|
| 2 |
+
license: mit
|
| 3 |
+
datasets:
|
| 4 |
+
- colored-dye/concept500-contrastive
|
| 5 |
+
language:
|
| 6 |
+
- en
|
| 7 |
+
base_model:
|
| 8 |
+
- google/gemma-2-2b-it
|
| 9 |
+
- google/gemma-2-9b-it
|
| 10 |
+
- Qwen/Qwen2.5-32B-Instruct
|
| 11 |
+
tags:
|
| 12 |
+
- steering-vector
|
| 13 |
+
---
|
| 14 |
+
|
| 15 |
+
# Towards Steering without Sacrifice: Principled Training of Steering Vectors for Prompt-only Interventions
|
| 16 |
+
|
| 17 |
+
OpenReview: https://openreview.net/forum?id=AaT3liS5PE
|
| 18 |
+
|
| 19 |
+
Paper: https://arxiv.org/abs/2605.05983
|
| 20 |
+
|
| 21 |
+
Data: https://huggingface.co/datasets/colored-dye/concept500-contrastive
|
| 22 |
+
|
| 23 |
+
Setups:
|
| 24 |
+
- `2b_l10`: 10th layer of google/gemma-2-2b-it
|
| 25 |
+
- `9b_l20`: 20th layer of google/gemma-2-9b-it
|
| 26 |
+
- `q25_32b_l32`: 32nd layer of qwen/Qwen2.5-32B-Instruct
|
| 27 |
+
|
| 28 |
+
Directory structure:
|
| 29 |
+
|
| 30 |
+
```
|
| 31 |
+
.
|
| 32 |
+
βββ 2b_l10
|
| 33 |
+
βΒ Β βββ outputs_add_free
|
| 34 |
+
βΒ Β βββ all -- full-sequence intervention
|
| 35 |
+
βΒ Β βΒ Β βββ lang -- Lang. objective
|
| 36 |
+
βΒ Β β β βββ 0 -- concept 0
|
| 37 |
+
βΒ Β β β βββ 1 -- concept 1
|
| 38 |
+
βΒ Β β β ...
|
| 39 |
+
βΒ Β βΒ Β βββ simpo -- SimPO objective
|
| 40 |
+
βΒ Β βββ f2+l2 -- prompt-only intervention (2 prefix tokens, 2 suffix tokens)
|
| 41 |
+
βΒ Β βββ lang
|
| 42 |
+
βΒ Β βββ simpo
|
| 43 |
+
```
|
| 44 |
+
|
| 45 |
+
|