Skip to content

predibase/lora_bakeoff

Repository files navigation

Lora Land: 310 Fine-tuned LLMs that Rival GPT-4

img

Low Rank Adaptation (LoRA) has emerged as one of the most widely adopted methods for Parameter Efficient Fine-Tuning (PEFT) of Large Language Models (LLMs). LoRA reduces the number of trainable parameters and memory usage while achieving comparable performance to full fine-tuning. We aim to assess the viability of training and serving LLMs fine-tuned with LoRA in real-world applications. First, we measure the quality of LLMs fine-tuned with quantized low rank adapters across 10 base models and 31 tasks for a total of 310 models. We find that 4-bit LoRA fine-tuned models outperform base models by 34 points and GPT-4 by 10 points on average. Second, we investigate the most effective base models for fine-tuning and assess the correlative and predictive capacities of dataset complexity heuristics in forecasting the outcomes of fine-tuning. Finally, we evaluate the latency and concurrency capabilities of LoRAX, an open-source LLM serving framework that facilitates the deployment of multiple LoRA fine-tuned models on a single GPU using shared base model weights and dynamic adapter loading. We use LoRAX to develop LoRA Land, a web application that hosts 25 LoRA fine-tuned Mistral-7B LLMs on a single NVIDIA A100 GPU with 80GB memory. This demonstration highlights the quality and cost-effectiveness of employing multiple specialized LLMs over a single, general-purpose LLM.

Tasks

The preprocessing code, prompt templates, and splits for all of our experiments can be found in the datasets/ directory.

img

Prompt design

img

img

Base models

img

Training configuration template

img

Based on Ludwig.

Results

img

img

Meta-correlations betweendataset complexity and model quality

img

Evaluation Harness

To use this eval harness, please set your PREDIBASE_API_TOKEN in .env before running the harness.

In the tasks directory, there is a subdirectory named glue_sst2, and inside of that subdirectory, there is a file named metadata.yaml. This is an example of the task-based organizational structure that the harness relies on. To add another task to the directory, it must follow the same convention: tasks/{task_name}/metadata.yaml.

The metadata.yaml file includes metadata like:

  • data_path: the path to the relevant dataset
  • prompt_template: the prompt template
  • target_col: the target column
  • metric_name: the metric function to use
  • split_column: the name of the column that defines splits in the dataset (optional)
  • split_column_value:which split to use (optional)

In the eval/glue_sst2 directory, the run.sh script serves as an example script to follow for other evaluations. There are three components of the script:

  • prep_pbase_requests.py -- Generates the JSON payloads to the REST API
  • pbase_request_parallel_processor.py -- Calls the REST API to get the responses from an adapter
  • parse_responses.py -- Calculates a metric score over the responses from the adapter (metrics listed in metric_fns.py)

In particular, the task flag refers to the name of the subdirectory containing the relevant metadata.yaml file.

To run the run.sh script, you will need to:

  1. Train your own adapter on the provided dataset and switch out the adapter_id parameter in the script with the model repo name and model version of your fine-tuned adapter.
  2. Change the deployment_base_model parameter to the name of the deployment you would like to use
  3. Change the tenant_id parameter to your tenant ID.

These three parameters can be found in the Predibase UI under "Models", "Prompt", and "Settings", respectively.

LoRAX Benchmarking

See lorax_load_test.js.

Sample command:

k6 run --env CONCURRENT_REQUESTS=2 --env NUM_INPUT_WORDS_LOWER_BOUND=90 --env NUM_INPUT_WORDS_UPPER_BOUND=110 --env MAX_NEW_TOKENS_LOWER_BOUND=90 --env MAX_NEW_TOKENS_UPPER_BOUND=110 --env SERVING_GATEWAY=serving.app.predibase.com --env TENANT=fd6c79 --env DEPLOYMENT_NAME=llama-2-7b-chat --env AUTH_TOKEN=pb_jcN0OPMdWt-yrgIg0aBnTA load_test.js

Web application

LoRA Land

img

Citation

@misc{loraland2024,
    title = {LoRA Land: 310 Fine-tuned LLMs that Rival GPT-4, A Technical Report},
    url = {https://predibase.com/blog/lora-land-fine-tuned-open-source-llms-that-outperform-gpt-4},
    author = {Justin Zhao, Timothy Wang, Wael Abid, Geoffrey Angus, Arnav Garg, Jeffery Kinnison, Piero Molino, Travis Addair, Devvret Rishi},
    month = {April},
    year = {2024}
}

About

No description, website, or topics provided.

Resources

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published