When Fine-Tuning Makes Sense: A Developer's Guide

https://getkiln.ai/blog/why_fine_tune_LLM_models_and_how_to_get_started

by scosman • 1 year ago

157 59 1 year ago

simonw10• 1 year ago

This is a post by a vendor that sells fine-tuning tools.

Here's a suggestion: show me a demo!

For the last two years I've been desperately keen to see just one good interactive demo that lets me see a fine-tuned model clearly performing better (faster, cheaper, more accurate results) than the base model on a task that it has been fine-tuned for - combined with extremely detailed information on how it was fine-tuned - all of the training data that was used.

If you want to stand out among all of the companies selling fine-tuning services yet another "here's tasks that can benefit from fine-tuning" post is not the way to do it. Build a compelling demo!

scosman5• 1 year ago

We don't sell fine-tuning tools - we're an open tool for finding the best way of running your AI workload. We support evaluating/comparing a variety of methods: prompting, prompt generators (few shot, repairs), various models, and fine-tuning from 5 different providers.

The focus of the tool is that it lets you try them all, side by side, and easily evaluate the results. Fine-tuning is one tool in a tool chest, which often wins, but not always. You should use evals to pick the best option for you. This also sets you up to iterate (when you find bugs, want to change the product, or new models comes out).

Re:demo -- would you want a demo or detailed evals and open datasets (honest question)? Single-shot examples are hard to compare, but the benefits usually come out in evals at scale. I'm definitely open to making this. Open for suggestions on what would be the most helpful (format and use case).

It's all on Github and free: https://github.com/kiln-ai/kiln

simonw10• 1 year ago

I want a web page I can go to where I can type a prompt (give me a list of example prompts too) and see the result from the base model on one side and the result from the fine-tuned model on the other side.

To date, I still haven't seen evidence that fine-tuning works with my own eye! It's really frustrating.

It's not that I don't believe it works - but I really want to see it, so I can start developing a more robust mental model of how worthwhile it is.

It sounds to me like you might be in a great position to offer this.

dist-epoch4• 1 year ago

I've seen many YouTube videos claiming that fine tuning can significantly reduce costs or make a smaller model perform like a larger one.

Most of them were not from fine-tuning tools or model sellers.

how it was fine-tuned - all of the training data that was used

It's not that sophisticated. You just need a dataset of prompts and the expected answer. And obviously a way to score the results, so you can guide the fine tuning.

simonw10• 1 year ago

I've seen those same claims, in videos and articles all over the place.

Which is why it's so weird that I can't find a convincing live demo to see the results for myself!

cleverwebble• 1 year ago

I can't really show an interactive demo, but my team at my day job has been fine tuning OpenAI models since GPT-3.5 and fine tuning can drastically improves output quality & prompt adherence. Heck, we found you can reduce your prompt to very simple instructions, and encode the style guidelines via your fine tuning examples.

This really only works though if:

1) The task is limited to a relatively small domain (relatively small could probably be misnomer, as most LLMs are trying to solve every-problem-all-at-once. As long as you are having it specialize in a specific field even, FT can help you achieve superior results.) 2) You have high quality examples (you don't need a lot, maybe 200 at most) Quality is often better than quantity here.

Often, distillation is all you need. Eg, do some prompt engineering on a high quality model (GPT-4.1, Gemini-Pro, Claude, etc.) - generate a few hundred examples, optionally (ideally) check for correctness via evaluations, and then fine tune a smaller, cheaper model. The new fine tuned model will not perform as well at generalist tasks as before, but it will be much more accurate at your specific domain, which is what most businesses care about.

jcheng• 1 year ago

200 examples at most, really?? I have been led to believe that (tens of) thousands is more typical. If you can get excellent results with that few examples, that changes the equation a lot.

tuyguntn• 1 year ago

Here's a suggestion: show me a demo!

Yes, yes and yes again!

Also, please don't use GIFs in your demos! It's freaking me out, because the speed of your GIF playback doesn't match my information absorption speed and I can't pause, look closely, go back, I just need to wait the second loop of your GIF

elliotto2• 1 year ago

Chiming in here to say that I was tasked to implement a fine tuning method for my AI startup and I also couldn't find any actual implemented outputs. There are piles of tutorials and blog posts and extensive documentation on hugging face transformers about the tools provided to do this, but I was unable to find a single demonstration of 'here is the base model output' vs 'here is the fine tuned output'. Doesn't have to be online like you suggested, even a screenshot or text blob showing how the fine tuning affected it would be useful.

I am in a similar boat to you where I have developed a great sense for how the bots will respond to prompting and how much detail and context is required because I've been able to iterate and experiment with this. But have no mental model at all about how fine tuning is meant to perform.

dedicate• 1 year ago

Interesting points! I'm always curious, though – beyond the theoretical benefits, has anyone here actually found a super specific, almost niche use case where fine-tuning blew a general model out of the water in a way that wasn't just about slight accuracy bumps?

scosman5• 1 year ago

Yup! I'll have to write some of these up. I can probably do open datasets and evals too. If you have use cases you'd like to see let me know! Some quick examples (task specific performance):

- fine-tuning improved performance of Llama 70B from 3.62/5 to (worse than Gemma 2B) to 4.27/5 (better than GPT 4.1), as measured by evals

- Generating valid JSON improved from <1% success rate to >95% after tuning

You can also optimize for cost/speed. I often see a 4x speedup and reducing costs by 90%+, while matching task-specific quality.

jampekka• 1 year ago

Don't you get valid JSON success rate of 100% with constrained decoding with any model?

genatron• 1 year ago

As an example Genatron is made possible by fine-tuning in order to create entire applications that are valid. It's similar to the valid json example, where you want to teach specific concepts through examples to ensure the correct syntactic and semantic outputs.

dist-epoch4• 1 year ago

Fine tuning is also about reducing costs. If you can bake half the prompt in the model through fine tuning, this can halve the running costs.

ldqm3• 1 year ago

I found Kiln a few months ago while looking for a UI to help build a dataset for fine-tuning a model on Grapheme-to-Phoneme (G2P) conversion. I’ve contributed to the repo since.

In my G2P task, smaller models were splitting phonemes inconsistently, which broke downstream tasks and caused a lot of retries - and higher costs. I fine-tuned Gemini, GPT-4o-mini, and some LLaMA and Qwen models on Fireworks.ai using Kiln, and it actually helped reduce those inconsistencies

simianwords5• 1 year ago

Related: what is the best way to augment the model with new knowledge other than at runtime using RAG?

scosman5• 1 year ago

Context window + prompt caching if you can't use RAG. Can add a lot to long context models, and their needle in haystack metrics keep getting better.

Why can't you use RAG?

simianwords5• 1 year ago

you lose coherence across chunks of context size. i wish i could spend compute to pre-train on some knowledge.

ijk• 1 year ago

Depends on the definition of "knowledge"; there's a lot of factors that go into it. Some of the common approaches are continued/continual pretraining and model editing (https://arxiv.org/pdf/2502.12598).

* Models are bad at learning that A=B implies B=A, let alone more complicated relations; augmenting the dataset with multiple examples with different phrasing/perspectives is important (https://arxiv.org/abs/2404.00213). The frequency that a relation occurs in the dataset affects the results (https://arxiv.org/html/2504.09597v2).

* You have to be able to balance preserving existing knowledge against the new knowledge (https://arxiv.org/abs/2502.14502). There are techniques like making sure your data mix corresponds to the original training data, but new data is primed by existing data so it gets complicated (https://arxiv.org/abs/2504.09522).

* Curriculum training (a la Phi) can be quite effective for training knowledge into base models at the very least.

* Continued pretraining is much more difficult than most finetuning, though it is possible (https://unsloth.ai/blog/contpretraining).

* Model editing of individual facts is possible but tricky because everything is interconnected but the model isn't great at figuring out reciprocal relationships (https://arxiv.org/abs/2310.16218). There's been some slow progress, though I find that few people are aware that it is even possible, despite the progress that has been made (https://github.com/zjunlp/KnowledgeEditingPapers).

The keywords you want are knowledge injection, domain adaptation, continual pretraining, model editing.

simianwords5• 1 year ago

This is exactly what I was talking about. I wonder why no one has tried to inject a critical code repository (at least 1 million LOC) and compare to common RAG methods?

The ones you have shown here are nice and simple like world cup statistics. Maybe we are nowhere near solving such complicated scenarios?

simonw10• 1 year ago

"What is the best way to augment the model with new knowledge other than at runtime using RAG?

I'm afraid the answer is "at runtime using RAG".

Don't fall into the trap of assuming that RAG has to mean janky vector embeddings though. There are many different ways to implement RAG. Good old fashioned FTS search (using tools like Elasticsearch or Solar or even PostgreSQL/MySQL/SQLite FTS) it's a lot less complicated and less expensive to set up and can provide extremely good results.

A lot of of the common RAG techniques were put together a couple of years ago when models were less capable and input limits were still around 8000 tokens.

The models today are much cheaper, far better and mostly have 100,000+ token input limits. This opens up all sorts of new RAG possibilities.

I am very excited at the moment by tool-driven RAG: implement a "search" tool for an LLM to use and prompt it to try several iterations on its search terms before it gives up.

o3 and o4-mini do this in ChatGPT with their web search tool and the results are extremely convincing.

simianwords5• 1 year ago

I agree that RAG does not have to be embeddings, RAG to me simply means augmenting new knowledge at run time no matter the method.

I would like to convince you that RAG may not be ideal and is simply an approximation of real learned data. RAG is inherently constrained by context length which means any understanding has to happen within chunks of size 100k tokens (as you pointed out). Keep in mind that you still lose high level semantic understanding as you increase the prompt token length to 100k even if needle in the haystack type of problems are solved at this level.

RAG introduces a severe limitation in understanding higher level semantic understanding across chunks. For instance, imagine a global variable shared across many modules causing some race conditions. This is extremely hard for RAG because it has to put many random modules in its context to deeply understand how the race condition happens. (to convince myself I must show that the linux codebase benefits from being indexed by an LLM and can solve hard to debug race conditions)

Another situation where RAG fails is where the you don't even know what to put in your context to get the answer. Imagine a prompt like "tell me two movies released in 2025 that are surprisingly similar in terms of story line". Maybe O3 can solve this particular problem but imagine I start adding more constraints?

ramoz• 1 year ago

There really isn't a good tool-calling model in open source, and I don't think the problem is fine-tuning.

jayavanth• 1 year ago

The best ones so far are fine-tunes. But I agree those numbers aren't great and we haven't figured out tool-calling yet

https://gorilla.cs.berkeley.edu/leaderboard.html

dist-epoch4• 1 year ago

Qwen3, Gemma, Mistral are open source and good at tool calling.

briian• 1 year ago

I think fine tuning is one of the things that makes verticalised agents so much better than general ones atm.

If agents aren’t specialised then every time they do anything, they have to figure out what to do and they don’t know what data matters, so often just slap entire web pages into their context. General agents use loads of tokens because of this. Vertical agents often have hard coded steps, know what data matters and already know what APIs they’re going to call. They’re far more efficient so will burn less cash.

This also improves the accuracy and quality.

I don't think this effect is as small as people say, especially when combined with the UX and domain specific workflows that verticalised agents allow for.

triyambakam2• 1 year ago

I have not yet heard of vertical agents. Any good resources?

simonw10• 1 year ago

I'm still fuzzy on what people mean when they say "agents".

mettamage• 1 year ago

Naive question, are there good tutorials/places that teach us to implement RAG and fine tune a model? I don't know if it's even feasible. At the moment I create AI workflows for the company I work at to (semi-)automate certain things. But it's not like I could fine-tune Claude. I'd need my own model for that. But would I need a whole GPU cluster? Or could it be done more easily.

And what about RAG? Is it hard to create embeddings?

I'm fairly new with the AI part of it all. I'm just using full-stack dev skills and some well written prompts.

scosman5• 1 year ago

Lot's of tools for each of those separately (RAG and fine-tuning). We're working on combining them but it's not ready yet.

You don't need a big GPU cluster. Fine-tuning is quite accessible via both APIs and local tools. It can be as simple as making API calls or using a UI. Some suggestions:

- https://getkiln.ai (my tool): let's you try all of the below, and compare/eval the resulting models

- API based tuning for closed models: OpenAI, Google Gemini

- API based tuning for open models: Together.ai, Fireworks.ai

- Local tuning for open models: https://unsloth.ai (can be run on Google Collab instances if you don't have local Nvidia GPUs).

Usually the building the training set and evaluating the resulting model is the hardest part. Another plug: Kiln support synthetic data gen and evals for these parts.

kaushalvivek• 1 year ago

Without concrete examples, this reads like an advertisement.

I am personlly very bullish on post-traning and fine-tuning. This artice doesn't do justice to the promise.

storus• 1 year ago

I thought that fine-tuning is no longer being done in the industry, instead transformer adapters like LoRA are being used? Having 1000 fine-tune models for each customer seems too heavy when one can have instead 1000 transformer adapters and swap them during the inference for each batch.

I mean there are tricks like Q-GaLore that allow training LLaMA-7B on a single 16GB GPU but LoRA still seems to be better for production to me.

nahnahno• 1 year ago

LoRA and QLoRA are still fine tuning I thought? Just updating a subset of parameters. You are still training a base model that was pre-trained (and possibly fine tuned after).

When Fine-Tuning Makes Sense: A Developer's Guide

Related Stories