explain what each chain tests and the three-chain design logic
Browse files
README.md
CHANGED
|
@@ -1,3 +1,26 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
# SAE × RL: Qwen2.5-0.5B on GSM8k (multi-condition)
|
| 2 |
|
| 3 |
Warm-start TopK SAEs trained on residual-stream activations of
|
|
|
|
| 1 |
+
---
|
| 2 |
+
license: apache-2.0
|
| 3 |
+
library_name: pytorch
|
| 4 |
+
language:
|
| 5 |
+
- en
|
| 6 |
+
pipeline_tag: feature-extraction
|
| 7 |
+
base_model: Qwen/Qwen2.5-0.5B-Instruct
|
| 8 |
+
datasets:
|
| 9 |
+
- openai/gsm8k
|
| 10 |
+
tags:
|
| 11 |
+
- sparse-autoencoder
|
| 12 |
+
- sae
|
| 13 |
+
- topk-sae
|
| 14 |
+
- interpretability
|
| 15 |
+
- mechanistic-interpretability
|
| 16 |
+
- ppo
|
| 17 |
+
- rlhf
|
| 18 |
+
- reasoning
|
| 19 |
+
- training-dynamics
|
| 20 |
+
- qwen
|
| 21 |
+
- qwen2.5
|
| 22 |
+
---
|
| 23 |
+
|
| 24 |
# SAE × RL: Qwen2.5-0.5B on GSM8k (multi-condition)
|
| 25 |
|
| 26 |
Warm-start TopK SAEs trained on residual-stream activations of
|