Fine-tuning Guide

Introduction

Large Language Models (LLMs) offer powerful general capabilities, but often require fine-tuning to excel at specific tasks or understand domain-specific language. Fine-tuning adapts a trained model to a smaller, targeted dataset, enhancing its performance for your unique needs.

This guide provides a step-by-step walkthrough for fine-tuning models using the Together AI platform. We will cover everything from preparing your data to evaluating your fine-tuned model.

We will cover:

1. Dataset Preparation: Loading a standard dataset, transforming it into the required format for supervised fine-tuning on Together AI, and uploading your formatted dataset to Together AI Files.
2. Fine-tuning Job Launch: Configuring and initiating a fine-tuning job using the Together AI API.
3. Job Monitoring: Checking the status and progress of your fine-tuning job.
4. Inference: Using your newly fine-tuned model via the Together AI API for predictions.
5. Evaluation: Comparing the performance of the fine-tuned model against the base model on a test set.

By following this guide, you'll gain practical experience in creating specialized LLMs tailored to your specific requirements using Together AI.

📘

Fine-tuning Guide Notebook

Here is a runnable notebook version of this fine-tuning guide: Fine-tuning Guide Notebook

Table of Contents

1. What is Fine-tuning?
2. Getting Started
3. Dataset Preparation
4. Starting a Fine-tuning Job
5. Monitoring Your Fine-tuning Job
6. Using Your Fine-tuned Model
7. Evaluating Your Fine-tuned Model
8. Advanced Topics

What is Fine-tuning?

Fine-tuning is the process of improving an existing LLM for a specific task or domain. You can enhance an LLM by providing labeled examples for a particular task which it can learn from. These examples can come from public datasets or private data specific to your organization.

Together AI facilitates every step of the fine-tuning process, from data preparation to model deployment. Together supports two types of fine-tuning:

1. LoRA (Low-Rank Adaptation) fine-tuning: Fine-tunes only a small subset of weights compared to full fine-tuning. This is faster, requires less computational resources, and is recommended for most use cases. Our fine-tuning API defaults to LoRA.

2. Full fine-tuning: Updates all weights in the model, which requires more computational resources but may provide better results for certain tasks.

Getting Started

Prerequisites

1. Register for an account: Sign up at Together AI to get an API key. New accounts come with $1 credit to get started.

2. Set up your API key:

   Shell

   export TOGETHER_API_KEY=your_api_key_here
   

3. Install the required libraries:

   Shell

   # Python
   pip install -U together datasets transformers tqdm
   

Choosing Your Model

The first step in fine-tuning is choosing which LLM to use as the starting point for your custom model:

• Base models are trained on a wide variety of texts, making their predictions broad
• Instruct models are trained on instruction-response pairs, making them better for specific tasks

For beginners, we recommend an instruction-tuned model:

• meta-llama/Meta-Llama-3.1-8B-Instruct-Reference is great for simpler tasks
• meta-llama/Meta-Llama-3.1-70B-Instruct-Reference is better for more complex datasets and domains

You can find all available models on the Together API here.

Dataset Preparation

Fine-tuning requires data formatted in a specific way. We'll use a conversational dataset as an example - here the goal is to improve the model on multi-turn conversations.

Data Formats

Together AI supports several data formats:

1. Conversational data: A JSON object per line, where each object contains a list of conversation turns under the "messages" key. Each message must have a "role" (system, user, or assistant) and "content". See details here.

   JSON

   {"messages": [{"role": "system", "content": "You are a helpful assistant."}, 
                 {"role": "user", "content": "Hello!"}, 
                 {"role": "assistant", "content": "Hi! How can I help you?"}]}
   

2. Instruction data: For instruction-based tasks with prompt-completion pairs. See details here.

3. Preference data: For preference-based fine-tuning. See details here.

4. Generic text data: For simple text completion tasks. See details here.

File Formats

Together AI supports two file formats:

1. JSONL: Simpler and works for most cases.
2. Parquet: Stores pre-tokenized data, provides flexibility to specify custom attention mask and labels (loss masking).

By default, it's easier to use JSONL. However, Parquet can be useful if you need custom tokenization or specific loss masking.

Example: Preparing the CoQA Dataset

Here's an example of transforming the CoQA dataset into the required chat format:

from datasets import load_dataset

# Load the dataset
coqa_dataset = load_dataset("stanfordnlp/coqa")

# The system prompt, if present, must always be at the beginning
system_prompt = "Read the story and extract answers for the questions.\nStory: {}"

def map_fields(row):
    # Create system prompt
    messages = [{"role": "system", "content": system_prompt.format(row["story"])}]
    
    # Add user and assistant messages
    for q, a in zip(row["questions"], row["answers"]["input_text"]):
        messages.append({"role": "user", "content": q})
        messages.append({"role": "assistant", "content": a})
    
    return {"messages": messages}
  
# Transform the data using the mapping function
train_messages = coqa_dataset["train"].map(map_fields, remove_columns=coqa_dataset["train"].column_names)

# Save to data to JSON file
train_messages.to_json("coqa_prepared_train.jsonl")

Loss Masking

In some cases, you may want to fine-tune a model to focus on predicting only a specific part of the prompt:

1. When using Conversational or Instruction Data Formats, you can specify train_on_inputs (bool or 'auto') - whether to mask the user messages in conversational data or prompts in instruction data.
2. For Conversational format, you can mask specific messages by assigning weights.
3. With pre-tokenized datasets (Parquet), you can provide custom labels to mask specific tokens by setting their label to -100.

Checking and Uploading Your Data

Once your data is prepared, verify it's correctly formatted and upload it to Together AI:

# Using Python
from together import Together
import os
import json

TOGETHER_API_KEY = os.getenv("TOGETHER_API_KEY")
WANDB_API_KEY = os.getenv("WANDB_API_KEY")  # Optional, for logging fine-tuning to wandb

# Check the file format
from together.utils import check_file

client = Together(api_key=TOGETHER_API_KEY)

sft_report = check_file("coqa_prepared_train.jsonl")
print(json.dumps(sft_report, indent=2))

assert sft_report["is_check_passed"] == True

# Upload the data to Together
train_file_resp = client.files.upload("coqa_prepared_train.jsonl", check=True)
print(train_file_resp.id)  # Save this ID for starting your fine-tuning job

# Using CLI
together files check "coqa_prepared_train.jsonl"
together files upload "coqa_prepared_train.jsonl"

The output from checking the file should look similar to:

{
  "is_check_passed": true,
  "message": "Checks passed",
  "found": true,
  "file_size": 23777505,
  "utf8": true,
  "line_type": true,
  "text_field": true,
  "key_value": true,
  "has_min_samples": true,
  "num_samples": 7199,
  "load_json": true,
  "filetype": "jsonl"
}

Starting a Fine-tuning Job

With our data uploaded, we can now launch the fine-tuning job using client.fine_tuning.create().

Key Parameters

• model: The base model you want to fine-tune (e.g., 'meta-llama/Meta-Llama-3.1-8B-Instruct-Reference')
• training_file: The ID of your uploaded training JSONL file
• validation_file: Optional ID of validation file (highly recommended for monitoring)
• suffix: A custom string added to create your unique model name (e.g., 'test1_8b')
• n_epochs: Number of times the model sees the entire dataset
• n_checkpoints: Number of checkpoints to save during training (for resuming or selecting the best model)
• learning_rate: Controls how much model weights are updated
• batch_size: Number of examples processed per iteration (default: "max")
• lora: Set to True for LoRA fine-tuning
• train_on_inputs: Whether to mask user messages or prompts (can be bool or 'auto')
• warmup_ratio: Ratio of steps for warmup

🔗 For an exhaustive list of all the available fine-tuning parameters refer to the Together AI Fine-tuning API Reference docs.

LoRA Fine-tuning (Recommended)

# Using Python - This fine-tuning job should take ~10-15 minutes to complete
ft_resp = client.fine_tuning.create(
    training_file = train_file_resp.id,
    model = 'meta-llama/Meta-Llama-3.1-8B-Instruct-Reference',
    train_on_inputs = "auto",
    n_epochs = 3,
    n_checkpoints = 1,
    wandb_api_key = WANDB_API_KEY,  # Optional, for visualization
    lora = True,   # Default True
    warmup_ratio = 0,
    learning_rate = 1e-5,
    suffix = 'test1_8b',
)

print(ft_resp.id)  # Save this job ID for monitoring

# Using CLI
together fine-tuning create \
  --training-file "file-id-from-upload" \
  --model "meta-llama/Meta-Llama-3.1-8B-Instruct-Reference" \
  --train-on-inputs auto \
  --lora \
  --n-epochs 3 \
  --n-checkpoints 1 \
  --warmup-ratio 0 \
  --learning-rate 1e-5 \
  --suffix "test1_8b" \
  --wandb-api-key $WANDB_API_KEY  # Optional

Full Fine-tuning

For full fine-tuning, simply omit the lora parameter:

# Using Python
ft_resp = client.fine_tuning.create(
    training_file = train_file_resp.id,
    model = 'meta-llama/Meta-Llama-3.1-8B-Instruct-Reference',
    train_on_inputs = "auto",
    n_epochs = 3,
    n_checkpoints = 1,
    warmup_ratio = 0,
    lora = False,  # Must be specified as False, defaults to True
    learning_rate = 1e-5,
    suffix = 'test1_8b_full_finetune',
)

# Using CLI
together fine-tuning create \
  --training-file "file-id-from-upload" \
  --model "meta-llama/Meta-Llama-3.1-8B-Instruct-Reference" \
  --train-on-inputs auto \
  --n-epochs 3 \
  --n-checkpoints 1 \
  --warmup-ratio 0 \
  --no-lora \
  --learning-rate 1e-5 \
  --suffix "test1_8b_full_finetune"

The response will include your job ID, which you'll use to monitor progress:

ft-d1522ffb-8f3e #fine-tuning job id

Monitoring Your Fine-tuning Job

Fine-tuning can take time depending on the model size, dataset size, and hyperparameters. Your job will progress through several states: Pending, Queued, Running, Uploading, and Completed.

You can monitor and manage the job's progress using the following methods:

• List all jobs: client.fine_tuning.list()
• Status of a job: client.fine_tuning.retrieve(id=ft_resp.id)
• List all events for a job: client.fine_tuning.list_events(id=ft_resp.id) - Retrieves logs and events generated during the job
• Cancel job: client.fine_tuning.cancel(id=ft_resp.id)
• Download fine-tuned model: client.fine_tuning.download(id=ft_resp.id)

Once the job is complete (status == 'completed'), the response from retrieve will contain the name of your newly created fine-tuned model. It follows the pattern: <your-account>/<base-model-name>:<suffix>:<job-id>.

Check Status via API

# Check status of the job
resp = client.fine_tuning.retrieve(ft_resp.id)
print(resp.status)

# This loop will print the logs of the job thus far
for event in client.fine_tuning.list_events(id=ft_resp.id).data:
    print(event.message)

# Using CLI
together fine-tuning retrieve "your-job-id"

Example output:

Fine tune request created
Job started at Thu Apr  3 03:19:46 UTC 2025
Model data downloaded for togethercomputer/Meta-Llama-3.1-8B-Instruct-Reference__TOG__FT at Thu Apr  3 03:19:48 UTC 2025
Data downloaded for togethercomputer/Meta-Llama-3.1-8B-Instruct-Reference__TOG__FT at $2025-04-03T03:19:55.595750
WandB run initialized.
Training started for model togethercomputer/Meta-Llama-3.1-8B-Instruct-Reference__TOG__FT
Epoch completed, at step 24
Epoch completed, at step 48
Epoch completed, at step 72
Training completed for togethercomputer/Meta-Llama-3.1-8B-Instruct-Reference__TOG__FT at Thu Apr  3 03:27:55 UTC 2025
Uploading output model
Compressing output model
Model compression complete
Model upload complete
Job finished at Thu Apr  3 03:31:33 UTC 2025

Dashboard Monitoring

You can also monitor your job on the Together AI jobs dashboard.

If you provided a Weights & Biases API key, you can view detailed training metrics on the W&B platform, including loss curves and more.

Using Your Fine-tuned Model

Once your fine-tuning job completes, your model will be available for use:

Option 1: Serverless LoRA Inference

If you used LoRA fine-tuning and the model supports serverless LoRA inference, you can immediately use your model without deployment. We can call it just like any other model on the Together AI platform, by providing the unique fine-tuned model output_name from our fine-tuning job.

🔗 See the list of all models that support LoRA Inference.

# The first time you run this it'll take longer to load the adapter weights for the first time
finetuned_model = ft_resp.output_name  # From your fine-tuning job response

user_prompt = "What is the capital of France?"

response = client.chat.completions.create(
    model = finetuned_model,
    messages=[
        {
            "role": "user",
            "content": user_prompt,
        }
    ],
    max_tokens=124,
)

print(response.choices[0].message.content)

You can also prompt the model in the Together AI playground by going to your models dashboard and clicking "OPEN IN PLAYGROUND". Read more about Serverless LoRA Inference here

Option 2: Deploy a Dedicated Endpoint

Another way to run your fine-tuned model is to deploy it on a custom dedicated endpoint:

1. Visit your models dashboard

2. Click "+ CREATE DEDICATED ENDPOINT" for your fine-tuned model

3. 

   Select hardware configuration and scaling options, including min and max replicas which affects the maximum QPS the deployment can support and then click "DEPLOY"

You can also deploy programmatically:

response = client.endpoints.create(
    display_name="Fine-tuned Meta Llama 3.1 8B Instruct 04-09-25",
    model="zainhas/Meta-Llama-3.1-8B-Instruct-Reference-test1_8b-e5a0fb5d",
    hardware="4x_nvidia_h100_80gb_sxm",
    autoscaling={
        "min_replicas": 1,
        "max_replicas": 1
    }
)

print(response)

⚠️ If you run this code it will deploy a dedicated endpoint for you. For detailed documentation around how to deploy, delete and modify endpoints see the Endpoints API Reference.

Once deployed, you can query the endpoint:

response = client.chat.completions.create(
    model="zainhas/Meta-Llama-3.1-8B-Instruct-Reference-test1_8b-e5a0fb5d-ded38e09",
    messages=[{
        "role": "user", 
        "content": "What is the capital of France?"
    }],
    max_tokens=128,
)

print(response.choices[0].message.content)

Evaluating Your Fine-tuned Model

To assess the impact of fine-tuning, we can compare the responses of our fine-tuned model with the original base model on the same prompts in our test set. This provides a way to measure improvements after fine-tuning.

Using a Validation Set During Training

You can provide a validation set when starting your fine-tuning job:

response = client.fine_tuning.create(
  training_file = 'your-training-file-id',
  validation_file = 'your-validation-file-id',
  n_evals = 10,  # Number of times to evaluate on validation set
  model = 'meta-llama/Meta-Llama-3.1-8B-Instruct-Reference',
)

Post-Training Evaluation Example

Here's a comprehensive example of evaluating models after fine-tuning, using the CoQA dataset:

1. First, load a portion of the validation dataset:

coqa_dataset_validation = load_dataset("stanfordnlp/coqa", split="validation[:50]")

2. Define a function to generate answers from both models:

from tqdm.auto import tqdm
from multiprocessing.pool import ThreadPool

base_model = "meta-llama/Meta-Llama-3.1-8B-Instruct-Turbo"  # Original model
finetuned_model = ft_resp.output_name  # Fine-tuned model

def get_model_answers(model_name):
    """
    Generate model answers for a given model name using a dataset of questions and answers.
    Args:
        model_name (str): The name of the model to use for generating answers.
    Returns:
        list: A list of lists, where each inner list contains the answers generated by the model.
    """
    model_answers = []
    system_prompt = "Read the story and extract answers for the questions.\nStory: {}"
    
    def get_answers(data):
        answers = []
        messages = [
            {
                "role": "system",
                "content": system_prompt.format(data["story"]),
            }
        ]
        for q, true_answer in zip(data["questions"], data["answers"]["input_text"]):
            try:
                messages.append(
                    {
                        "role": "user",
                        "content": q
                    }
                )
                response = client.chat.completions.create(
                    messages=messages,
                    model=model_name,
                    max_tokens=64
                )
                answer = response.choices[0].message.content
                answers.append(answer)
            except Exception:
                answers.append("Invalid Response")
        return answers

    # We'll use 8 threads to generate answers faster in parallel
    with ThreadPool(8) as pool:
        for answers in tqdm(pool.imap(get_answers, coqa_dataset_validation), total=len(coqa_dataset_validation)):
            model_answers.append(answers)

    return model_answers

3. Generate answers from both models:

base_answers = get_model_answers(base_model)
finetuned_answers = get_model_answers(finetuned_model)

4. Define a function to calculate evaluation metrics:

import transformers.data.metrics.squad_metrics as squad_metrics

def get_metrics(pred_answers):
    """
    Calculate the Exact Match (EM) and F1 metrics for predicted answers.
    Args:
        pred_answers (list): A list of predicted answers.
    Returns:
        tuple: A tuple containing EM score and F1 score.
    """
    em_metrics = []
    f1_metrics = []

    for pred, data in tqdm(zip(pred_answers, coqa_dataset_validation), total=len(pred_answers)):
        for pred_answer, true_answer in zip(pred, data["answers"]["input_text"]):
            em_metrics.append(squad_metrics.compute_exact(true_answer, pred_answer))
            f1_metrics.append(squad_metrics.compute_f1(true_answer, pred_answer))

    return sum(em_metrics) / len(em_metrics), sum(f1_metrics) / len(f1_metrics)

5. Calculate and compare metrics:

# Calculate metrics for both models
em_base, f1_base = get_metrics(base_answers)
em_ft, f1_ft = get_metrics(finetuned_answers)

print(f"Base Model - EM: {em_base:.2f}, F1: {f1_base:.2f}")
print(f"Fine-tuned Model - EM: {em_ft:.2f}, F1: {f1_ft:.2f}")

You should get figures similar to the table below:

We can see that the fine-tuned model performs significantly better on the test set, with a large improvement in both Exact Match and F1 scores.

Advanced Topics

Continuing a Fine-tuning Job

You can continue training from a previous fine-tuning job:

response = client.fine_tuning.create(
  training_file = 'your-new-training-file-id',
  from_checkpoint = 'previous-finetune-job-id',
  wandb_api_key = 'your-wandb-api-key',
)

together fine-tuning create \
  --training-file "your-new-training-file-id" \
  --from-checkpoint "previous-finetune-job-id" \
  --wandb-api-key $WANDB_API_KEY

You can specify a checkpoint by using:

• The output model name from the previous job
• Fine-tuning job ID
• A specific checkpoint step with the format ft-...:{STEP_NUM}

To check all available checkpoints for a job, use:

together fine-tuning list-checkpoints {FT_JOB_ID}

Training and Validation Split

To split your dataset into training and validation sets:

split_ratio=0.9  # Specify the split ratio for your training set

total_lines=$(wc -l < "your-datafile.jsonl")
split_lines=$((total_lines * split_ratio))

head -n $split_lines "your-datafile.jsonl" > "your-datafile-train.jsonl"
tail -n +$((split_lines + 1)) "your-datafile.jsonl" > "your-datafile-validation.jsonl"

Using a Validation Set During Training

A validation set is a held-out dataset to evaluate your model performance during training on unseen data. Using a validation set provides multiple benefits such as monitoring for overfitting and helping with hyperparameter tuning.

To use a validation set, provide validation_file and set n_evals to a number above 0:

response = client.fine_tuning.create(
  training_file = 'your-training-file-id',
  validation_file = 'your-validation-file-id',
  n_evals = 10,  # Number of evaluations over the entire job
  model = 'meta-llama/Meta-Llama-3.1-8B-Instruct-Reference',
)

At set intervals during training, the model will be evaluated on your validation set, and the evaluation loss will be recorded in your job event log. If you provide a W&B API key, you'll also be able to see these losses in the W&B dashboard.

Conclusion

Fine-tuning LLMs with Together AI allows you to create specialized models tailored to your specific requirements. By following this guide, you've learned how to:

1. Prepare and format your data for fine-tuning
2. Launch a fine-tuning job with appropriate parameters
3. Monitor the progress of your fine-tuning job
4. Use your fine-tuned model via API or dedicated endpoints
5. Evaluate your model's performance improvements
6. Work with advanced features like continued training and validation sets