Skip to content

OpenAssistant

Finetune a model with OpenAssistant dataset

See: OpenAssistant Dataset

For more detailed instructions on finetuning, refer to Samantha

The configurations are still pretty raw and could use more tuning, but should work.

Usage

# Basic Usage for 7B model on 1 GPU
forgather -t 7b_1gpu.yaml train -M /path/to/model

# Note: The "pipeline" configs will only work with Forgather
# pipeline compatible models. See Samantha for instructions on model conversion.

# 7B model on 2 GPU pipeline
forgather -t 7b_2gpu.yaml train -M /path/to/model

# 7B model on 4 GPU pipeline
forgather -t 7b_4gpu.yaml train -M /path/to/model

Test Trained Model Performance

Measure loss/perplexity on the OpenAssistant test split with forgather eval:

forgather eval test openassistant -M /path/to/model --dtype bfloat16

See docs/guides/evaluating-models.md for the full workflow (trainer choices, multi-GPU, pipeline parallelism, ...) and forgather eval list for the other shipped configs.