Hi everyone, we’ve been working on benchmarking different open-source LLMs. We measure, in particular, on the performance of these models once finetued (via QLoRA) on classic NLP downstream tasks like summarization and classification. We also put particular emphasis on benchmarking inference time/cost for these models once deployed.

We’ve just ran our study on the new Zephyr-7B-beta model, a DPO-tuned version of Mistral-7B.

We first tested out-of-the-box performance of Zephyr for summarization under zero-shot and few-shot (for classification, we couldn’t do few-shot because of context length, and we haven’t tried zero-shot since most other open source models gave subpar results).

Then we tested the performance after QLoRA fine-tuning and saw substantial performance boost (as it should). Afterwards we experimented levers we can pull to increase model performance (using NEFTune and/or finetuning on all modules as opposed to attention modules only).

Summarization

Dataset Used: Newsgroup

Rank: 64

• We see that zero-shot and few-shot performance is already pretty good out-of-the box
• QLoRA was able to refine the syntactic style and pithiness of outputs to match that of the training set
• NEFTune did not improve summarization performance noticeably
• Tuning all modules (as opposed to attention modules) yielded slightly better results

Classification

Dataset Used: Samsum

Rank: 8

• NEFTune boosted performance in low-data regimes
• Tuning all modules has achived ~10x sample efficiency and better better performance at 100% traning fraction