This model was trained as part of a series of experiments testing the performance of pure DPO vs SFT vs ORPO, all supported by Unsloth/Huggingface TRL.
Note: Extremely buggy, not recommended for use. However, it didn't massively overfit like #3, so it could be usable still.
The training was somewhat unstable, so the optimal bound for LR seems to be around [1e-5, 1e-4].
Benchmarks
For some reason the OpenLLM leaderboard refuses to bench this model, so I guess we will never know how well it performs.
Training Details
Duration: ~10-12 hours on one Kaggle T4 with Unsloth
Model: https://huggingface.co/unsloth/mistral-7b-v0.2-bnb-4bit
Dataset: https://huggingface.co/datasets/argilla/dpo-mix-7k
Rank: 8
Alpha: 16
Learning rate: 1e-4
Beta: 0.1
Batch size: 8
Epochs: 1
Learning rate scheduler: Linear
Prompt Format: ChatML
<|im_start|>system
You are a helpful assistant.<|im_end|>
<|im_start|>user
Why is the sky blue?<|im_end|>
<|im_start|>assistant
WanDB Reports
