OhhMoo/sae-rl-qwen05b-strict-activations
SAE × RL: Qwen2.5-0.5B on GSM8k Warm-start TopK SAEs trained on 8 PPO checkpoints × 4 residual-stream layers of Qwen2.5-0.5B-Instruct fine-tuned on GSM8k. Purpose: measure what RL changes inside the model. What I did PPO on Qwen2.5-0.5B-Instruct on GSM8k with a flexible reward (last number in the response = correct). 2 GPUs, verl FSDP + vLLM rollout. Checkpointed every 10 steps; ran to step 200. Collected residual-stream activations at layers {6, 12, 18, 23} from… See the full description on the dataset page: https://huggingface.co/datasets/OhhMoo/sae-rl-qwen05b-strict-activations.
| Author / organization | OhhMoo |
|---|---|
| License (as declared) | apache-2.0 |
| Downloads | 183 (as reported by Hugging Face) |
| Likes | 0 |
| Created | 1 May 2026 |
| Last updated at source | 1 May 2026 |
Popularity, last 2 days
183 now. Counts as reported by the source, snapshotted daily by Silytics.
Recent changes
- 16 Oct 2026 Description: SAE × RL: Qwen2.5-0.5B on GSM8k Warm-start TopK SAEs trained on 8 PPO checkpoints × 4 residual-stream layers of Qwen2.5-0.5B-Instruct fine-tuned on GSM8k. Purpose: measure what RL changes inside the model. What I did PPO on Qwen2.5-0.5B-Ins → SAE × RL: Qwen2.5-0.5B on GSM8k Warm-start TopK SAEs trained on 8 PPO checkpoints × 4 residual-stream layers of Qwen2.5-0.5B-Instruct fine-tuned on GSM8k. Purpose: measure what RL changes inside the model. What I did PPO on Qwen2.5-0.5B-Ins
Changes detected by Silytics between refreshes of the source.
Silytics shows public metadata from Hugging Face and links to the original. Information may be incomplete or out of date. Last refreshed 17 October 2026. Licenses and terms of the original apply to the work itself. Not financial, legal or professional advice.