databricks-ml-examples

databricks/databricks-ml-examples is a repository to show machine learning examples on Databricks platforms.

Currently this repository contains:

• llm-models/: Example notebooks to use different State of the art (SOTA) models on Databricks.

SOTA LLM examples

Databricks works with thousands of customers to build generative AI applications. While you can use Databricks to work with any generative AI model, including commercial and research, the table below lists our current model recommendations for popular use cases. Note: The table only lists open source models that are for free commercial use.

Use case	Quality-optimized	Balanced	Speed-optimized
Text generation following instructions	MPT-30B-Instruct Llama-2-70b-chat-hf	mistral-7b MPT-7B-Instruct MPT-7B-8k-Instruct Llama-2-7b-chat-hf Llama-2-13b-chat-hf
Text embeddings (English only)		bge-large-en-v1.5(0.3B) e5-large-v2 (0.3B)	bge-base-en-v1.5 (0.1B) e5-base-v2 (0.1B)
Transcription (speech to text)		whisper-large-v2(1.6B) whisper-medium (0.8B)
Image generation		stable-diffusion-xl
Code generation	CodeLlama-34b-hf CodeLlama-34b-Instruct-hf CodeLlama-34b-Python-hf (Python optimized)	CodeLlama-13b-hf CodeLlama-13b-Instruct-hf CodeLlama-13b-Python-hf (Python optimized) CodeLlama-7b-hf CodeLlama-7b-Instruct-hf CodeLlama-7b-Python-hf (Python optimized)

• To get a better performance on instructor-xl, you may follow the unified template to write instructions.

Model Evaluation Leaderboard

Text generation models

The model evaluation results presented below are measured by the Mosaic Eval Gauntlet framework. This framework comprises a series of tasks specifically designed to assess the performance of language models, including widely-adopted benchmarks such as MMLU, Big-Bench, HellaSwag, and more.

Model Name	Core Average	World Knowledge	Commonsense Reasoning	Language Understanding	Symbolic Problem Solving	Reading Comprehension
Mistral-7B-v0.1	0.522	0.558	0.513	0.555	0.342	0.641
falcon-40b	0.501	0.556	0.55	0.535	0.269	0.597
falcon-40b-instruct	0.5	0.542	0.571	0.544	0.264	0.58
Llama-2-13b-hf	0.479	0.515	0.482	0.52	0.279	0.597
Llama-2-13b-chat-hf	0.476	0.522	0.512	0.514	0.271	0.559
Mistral-7B-Instruct-v0.1	0.469	0.48	0.502	0.492	0.266	0.604
mpt-30b-instruct	0.465	0.48	0.513	0.494	0.238	0.599
mpt-30b	0.431	0.494	0.47	0.477	0.234	0.481
Llama-2-7b-chat-hf	0.42	0.476	0.447	0.478	0.221	0.478
Llama-2-7b-hf	0.401	0.457	0.41	0.454	0.217	0.465
mpt-7b-8k-instruct	0.36	0.363	0.41	0.405	0.165	0.458
mpt-7b-instruct	0.354	0.399	0.415	0.372	0.171	0.415
mpt-7b-8k	0.354	0.427	0.368	0.426	0.171	0.378
falcon-7b	0.335	0.371	0.421	0.37	0.159	0.355
mpt-7b	0.324	0.356	0.384	0.38	0.163	0.336
falcon-7b-instruct	0.307	0.34	0.372	0.333	0.108	0.38

Other examples:

• DIY LLM QA Bot Accelerator
• Biomedical Question Answering over Custom Datasets with LangChain and Llama 2 from Hugging Face
• DIY QA LLM BOT
• Tuning the Finetuning: An exploration of achieving success with QLoRA
• databricks-llm-fine-tuning