NVIDIA-NeMo / NVIDIA-NeMo/Automodel

Nemotron-Nano-9b-V2 has low full-weights SFT accuracy

Open
#893 0 comments 0 reactions 0 assignees View on GitHub

Nobody has claimed this yet.

bug
Dominant language
Python
Stars
963
Forks
318
Avg merge
3d 20h
Merged PRs (30d)
143

Description

Describe the bug

Nemotron-Nano-9b-V2 full weights SFT accuracy is really low. The rouge score is 0.07. However, the older AutoModel training scripts give a higher rouge score (0.23) for the same hyper-parameters.

Steps/Code to reproduce bug

torchrun --nproc-per-node=8 /opt/Automodel/examples/llm_finetune/finetune.py --config config.yaml

Config.yaml - config.yaml.rtf

Dataset is attached.

email-composition-train.tar.gz

Expected behavior

Base model rouge score is at least 0.23. Full weights SFT should give at least 0.23.

Additional context

Add any other context about the problem here.

Contributor guide

Open the contributing guide

First steps

  1. Read the whole issue, then the project's contributing guide.
  2. Comment on the issue to say you are picking it up — it saves two people doing the same work.
  3. Fork the repository and make your change on a branch.
  4. Open a pull request that references the issue number.

Research direction

Reproduce the issue with the provided torchrun command, config.yaml, and email-composition dataset, starting at examples/llm_finetune/finetune.py. Compare this path with the older AutoModel training scripts using the same hyper-parameters. Done means full-weights SFT reaches the expected ROUGE score of at least 0.23.

Written by the indexing model from the issue text.

Assessment

Tech stack
python, pytorch
Domain
machine-learning
Issue type
Bug
Difficulty
4/5
Estimated time
3-5 days
Activity status
Stale
Clarity
Mostly clear
Newbie friendliness
38/100

Get new issues in your inbox

A short digest of beginner-friendly GitHub issues.