I’ve been doing some fine-tuning experiments lately, and while fine-tuning commercial LLMs like GPT-3.5 is pretty easy and cheap, but there are restrictions on model sharing, control of the algorithm, and cost are a bit annoying. An alternative is to fine-tune open-weight models like Llama-3-8B, etc. It is relatively easy and cheap to do it on the cloud. Some notes on the ecosystem for fine-tuning your own LLMs…with an emphasis on doing this remotely for the GPU poor…

What do I care about?

  • Ease of use: I’m interested in doing science, rather than screwing around with a lot of fiddly VM configurations. Fine-tuning as a service is great! Chose my hyperparameters for me! Make it easy for me to evaluate on test items!
  • Clarity/Reproducibility: We want to report things in scholarly work, so a clear indication of chosen hyperparameters, etc. is important so others can follow us without having to rely on a particular commercial service (which might disappear…). Also valuable is some way of sharing trained models with others (e.g., by downloading weights).

An educational resource

  • In May/June 2024 there was a fun LLM-fine-tuning course; the videos are now available for free online. The content below are some highlights based on my own experiences with the various services/tools presented in the course. So if you want more details or examples on any of these, maybe check out that resource first.

(presented in no particular order)

OpenPipe

  • Ease of Use: High. I tried this for an example problem… it was pretty easy to use with the web-interface, and the use of an OpenAI-compatible data format and REST API made it easy to compare a few different models (incl. OpenAI, LLama-3 variants, Mistral-7b) in a consistent way.
  • Reproducibility: Medium. No ability to control (or see) what hyperparameters were used or other specific details. But easy to download final trained model weights.

OpenPipe is another service that is designed around being a pass-through that you can use to collect training data from GPT-4 and then use that stored data to fine-tune your own models in production. So there’s a way in which we are mis-using this service.

However is easy to upload your own data files; they use the OpenAI JSONL format for training data, but add a key split which lets you specify train/test split (otherwise it will randomly select 10%). But, (as of late-June 2024)there is no way to download these test results. They want you to do LLM-based reranking and evaluation as a comparison, but don’t support a simple data download. So you have to re-call the model to do inferences if you want this data (which is slightly wasteful, but so it is). Those evaluation capabilities, are mostly based on GPT-4 rescoring rather than the types of supervised checking that we use.

Current base models support a few different things, including Llama-3 variants and a pass-through to GPT-3.5. (update July 2024) They are now supporting the newest Llama3 variants and pass-through to GPT-4o-mini fine tuning.

One nice feature is an an “export weights” option on the model information page that lets you download them (assuming you are not using GPT-3.5). Great for reproducibility!

What is frustrating (circa late June/early July 2024) is that you can’t see what the automatically-chosen hyperparameters were. You’re really in the dark here. (The exception is for fine-tuning the OpenAI models, in which case you can go to platform.openai.com and find all the goodies there.)

You’ve got to use the web-interface; as of late-June 2024, the API for fine-tuning is still in beta.

Predibase

Predibase is a business exclusively oriented around open-source model fine-tuning using Lorax. A quick look at the docs suggests that they are very flexible about input data formats and they support a wide variety of models. Unclear to me how you can export/share your models, but give it a shot.

Under the hood, it uses Ludwig to configure fine-tuning and Lorax to serve the adapters efficiently. You can run these locally.

They also maintain a useful leaderboard (and check link there for their ArXiV paper) comparing fine-tuning performance on a variety of tasks. As of late-June 2024, llama-3-8b is pretty much the best choice, fwiw.

Fireworks

Fireworks is yet another open-LLM-as-a-service provider also making a play for finetuning.

A strength (circa late July 2024) compared to the services mentioned above, is that they appear to support grammar-based structured output

They also have a few fine-tuned models of their own, such as a Llama3 that does better function calling (claimed to be comparable to GPT-4o, but at 10% of the cost and 3x speed)

OpenAI

(24 July 2024) I got my start in fine-tuning with gpt-3.5-turbo and there’s something magical about how polished the experience is. Having now done a bunch of fine-tuning with Llama-3 and Mistral, it’s still easier to get good results on OpenAI—just let it pick hyperparameters and the results are pretty good.

Of course, the downside is that it is a black box. Another downside is that your fine-tuned model can only be used by people with an API key within your organization—in other words, you’ve got to pay the bill if you want external users to run the model. This is kind of a drag for disseminating models to the public. But on the other hand, they’re cheap enough to train that you can have someone else train them.

But the new gpt-4o-mini fine-tuning that has been just released is great. Super cheap to train and to do inference—about 50% to train and 80% cheaper on inference of gpt-3.5-turbo and preliminary experiments suggest it is slightly better (or at least not worse). This may have to do with the longer training context window 65k tokens (and that may also explain why it does better than some of our llama-3 fine tunes)

Huggingface / autotrain

Autotrain is a general purpose open-source software package for no-code training/fine-tuning models (LLMS, image generation, tabular data). It is created/maintained by Huggingface (gotta support our NYC homies).

You can run it locally, or hosted with compute provide by HF, run in a google collab, or inside a VM hosted on JarvisLab (for example); in that way it is quite flexible. The GUI has a basic and advanced mode, but you can export (or modify) the inputs in a JSON/YAML config file. Under the hood it is powered by the accelerate package (also maintained by HF), which handles distributing load over multiple GPUs, data packing, etc.

The integration with HF is quite convenient: You can pull data hosted on HF and pull it in, pull a base model from HF, and then push the resulting model to HF. It handles most of the common data formats (e.g., ChatML), and fine-tuning modes (e.g. qLORA, etc); not necessarily the most cutting edge package or the finest granularity of control over data, but covers most of the common options.

Axolotl / JarvisLabs

Axolotl is a python package that provides an abstraction layer for fine-tuning. Essentially everything gets driven by a long YAML file. You can run Axolotl anywhere you want, but it is convenient to have all the dependencies preconfigured.

JarvisLabs is a cloud compute service (GPU by the hour). Their prices seem OK; main advantage seems to be a variety of pre-built packages for common AI workflows. They provide a pre-built container for Axolotl fine-tuning which makes it pretty easy and the video they provide is a good starting point. A suggestion is to use tiny-llama/qlora.yaml as hello-world example instead, as it is faster than the example they provide.

Replicate

Replicate is a service that hosts open-weight models, lets you post them (by using Cog), and then serves them to public users, including a lightweight web-UI.

Additionally (in beta) it supports model fine-tuning, however as of 12 June 2024 their Current list of trainable language models only includes Llama-2-class models. The developer says it is best to use the chat versions; for now this does not include the models that support structured grammars

Axolotl / Modal

Modal provides way to do remote procedure calling in python made easy…especially focused on data-intensive applications where you want the data to live remotely near the compute. There’s a lot of slickness that extends beyond LLMs, but…

Their tutorial documentation has a demonstration for fine-tuning with Axolotl

This is a way to stay closer to the metal (you are doing Axolotl after all), but without all of the complexity of configuring an entire virtual machine instance. It’s also useful for doing numerical compute type things, but that’s a topic for another blog post.

Mistral

Mistral provides fine-tuning as a service through LaPlateforme but also has a python package to facilitate fine-tuning their models your own (VM) hardware.

Cloud: Minimum fee of $4 and monthly storage fee of $2 for each model. Models are apparently only usable within a billing organization (like OpenAI), which is a downside for sharing. As of late-June 2024, they only support fine tuning open-mistral-7b and mistral-small. There’s a limit of 512 MB training data and 1MB validation data. It’s pretty bare-bones

Local/VM: There’s a video on the internet walking through the process…doesn’t look too hard. You can do it on collab or a A100 instance. It’s pretty bare bones. The advantage of doing it this way is that it provides the ability to do batch inferences more efficiently (by just driving it directly in Python rather than having to make http requests.)

Mac stuff

I’ve been looking for an excuse to buy a maxed out Mac Studio with 192GB of RAM…

There’s a HF community for MLX users that provides basic examples of converting/running/fine-tuning LLMs.

Folks have also done fine-tuning of Mistral and Llama on M2Ultras and also at 475 tokens/sec; given that our past work was about 1M tokens of fine-tuning, we could crank it out in 35 minutes which ain’t bad.

Unsloth

Unsloth is a library for performance tunning LLMs and they have a worked example of fine tuning Llama-3 with UnSloth and Export to Ollama on a Collab instance

Miscellaneous tips

Other things seen on the internet