Hello, and thank you for releasing this excellent work!
I’m a graduate student currently attempting to reproduce the HaloScope paper results for research purposes (hallucination detection on LLMs).
While setting up the environment from the provided requirements.txt, I encountered several installation issues, mainly related to dependency conflicts and packages not found on PyPI.
Specifically:
apex==0.9 → returns No matching distribution found
Multiple nvidia-* packages (e.g., nvidia-cublas-cu12, nvidia-cudnn-cu12) cause CUDA version conflicts with existing PyTorch binaries.
Some versions (e.g., flash-attn, vllm-flash-attn) require manual compilation or are not available via pip.
Even though the README suggests installation via:
pip install -r requirements.txt
some of these packages cannot be resolved automatically through pip or conda.
Could you please clarify:
Whether the listed dependencies were tested on a specific CUDA / PyTorch version (e.g., CUDA 11.8 or 12.1)?
Is apex required for running the core hallucination detection pipeline (e.g., hal_det_llama.py), or is it optional?
Would you consider updating requirements.txt or providing a minimal environment list (only required packages for inference)?
Any guidance on reproducing your environment would be greatly appreciated!
Thank you very much for your time and for open-sourcing HaloScope — it’s a valuable contribution to the hallucination detection research community.
Looking forward to your advice.
Hello, and thank you for releasing this excellent work!
I’m a graduate student currently attempting to reproduce the HaloScope paper results for research purposes (hallucination detection on LLMs).
While setting up the environment from the provided requirements.txt, I encountered several installation issues, mainly related to dependency conflicts and packages not found on PyPI.
Specifically:
apex==0.9 → returns No matching distribution found
Multiple nvidia-* packages (e.g., nvidia-cublas-cu12, nvidia-cudnn-cu12) cause CUDA version conflicts with existing PyTorch binaries.
Some versions (e.g., flash-attn, vllm-flash-attn) require manual compilation or are not available via pip.
Even though the README suggests installation via:
pip install -r requirements.txt
some of these packages cannot be resolved automatically through pip or conda.
Could you please clarify:
Whether the listed dependencies were tested on a specific CUDA / PyTorch version (e.g., CUDA 11.8 or 12.1)?
Is apex required for running the core hallucination detection pipeline (e.g., hal_det_llama.py), or is it optional?
Would you consider updating requirements.txt or providing a minimal environment list (only required packages for inference)?
Any guidance on reproducing your environment would be greatly appreciated!
Thank you very much for your time and for open-sourcing HaloScope — it’s a valuable contribution to the hallucination detection research community.
Looking forward to your advice.