Reproducing results for “From Reasoning to Answer…” — Hugging Face changes?

[Ykid]

Hello!

Thank you for releasing the code for From Reasoning to Answer: Empirical, Attention-Based and Mechanistic Insights into Distilled DeepSeek R1 Models. I’m attempting to reproduce the experiments and ran into an issue that seem related to recent changes on the Hugging Face repo.

Could you advise on how best to address this? Thank you!

(r2a) a@a:~/x/ACV/misc/R2A$ python trace_collection.py \
    --model_path "deepseek-ai/DeepSeek-R1-Distill-Qwen-1.5B" \
    --dataset "WildBench" \
    --with_reasoning \
    --limit 1
Output path set to: output/reasoning_traces/WildBench_DeepSeek-R1-Distill-Qwen-1.5B_withR_sampling_0.jsonl
Loading model: deepseek-ai/DeepSeek-R1-Distill-Qwen-1.5B
Loading dataset: allenai/WildBench
README.md: 8.22kB [00:00, 26.2MB/s]
^CTraceback (most recent call last):
  File "/opt/conda/envs/r2a/lib/python3.11/site-packages/huggingface_hub/utils/_http.py", line 409, in hf_raise_for_status
    response.raise_for_status()
  File "/opt/conda/envs/r2a/lib/python3.11/site-packages/requests/models.py", line 1026, in raise_for_status
    raise HTTPError(http_error_msg, response=self)
requests.exceptions.HTTPError: 404 Client Error: Not Found for url: https://huggingface.co/datasets/allenai/WildBench/resolve/26c49eb39d7d5ce2099b0bbafed5a88dcce954ec/WildBench.py