Issue
The README mentions that OpenMath is designed to be reproducible on free Colab (T4) GPUs.
However, the repository does not currently include a requirements.txt (or equivalent)
to specify Python dependencies and versions.
Given that QLoRA fine-tuning and inference are highly sensitive to library versions
(e.g., transformers, peft, bitsandbytes, torch), this makes reproducibility difficult
outside of the original environment.
Impact
- Users may encounter version conflicts or silent failures
- Results may not be reproducible across environments
- Onboarding new contributors becomes harder
Suggested Improvement
Add a requirements.txt file specifying tested versions for:
- torch
- transformers
- peft
- bitsandbytes
- accelerate
- safetensors
- datasets (if used for evaluation)
Optionally, note the tested Python and CUDA versions in the README.
This would significantly improve reproducibility and ease of use.
Kindly Assign me this issue under OSCG ’26
Issue
The README mentions that OpenMath is designed to be reproducible on free Colab (T4) GPUs.
However, the repository does not currently include a
requirements.txt(or equivalent)to specify Python dependencies and versions.
Given that QLoRA fine-tuning and inference are highly sensitive to library versions
(e.g., transformers, peft, bitsandbytes, torch), this makes reproducibility difficult
outside of the original environment.
Impact
Suggested Improvement
Add a
requirements.txtfile specifying tested versions for:Optionally, note the tested Python and CUDA versions in the README.
This would significantly improve reproducibility and ease of use.
Kindly Assign me this issue under OSCG ’26