Finetuning demos for LLMs

This repo contains demos for finetuning of Large Language Models (LLMs), like Meta's llama-3. In particular, we focus on tuning for short-form instruction following capabilities and conversational chat capabilities.

---

📚 Background

The goal of instruction-tuning is to train LLMs that are capable of following natural language instructions to perform a wide range of tasks. The below was captured from the "State of GPTs" talk by Andrej Karpathy:

For more background, see any number of excellent papers on the subject, like InstructGPT (2022).

🔎 Finetuned models

See src for all finetuning runs. Here are some favorites:

1. dfurman/Llama-3-8B-Orpo-v0.1
2. dfurman/Llama-3-70B-Orpo-v0.1
3. dfurman/Llama-2-70B-Instruct-v0.1
   • Note: This model was ranked 6th on 🤗's Open LLM Leaderboard in Aug 2023
4. dfurman/Mixtral-8x7B-Instruct-v0.1
5. dfurman/Mistral-7B-Instruct-v0.2

💻 Usage

Note: Use the code below to get started. Be sure to have a GPU-enabled cluster.

Setup

!pip install -qU transformers accelerate

from transformers import AutoTokenizer
import transformers
import torch

if torch.cuda.get_device_capability()[0] >= 8:
    !pip install -qqq flash-attn
    attn_implementation = "flash_attention_2"
    torch_dtype = torch.bfloat16
else:
    attn_implementation = "eager"
    torch_dtype = torch.float16

model = "dfurman/Llama-3-8B-Orpo-v0.1"

tokenizer = AutoTokenizer.from_pretrained(model)
pipeline = transformers.pipeline(
    "text-generation",
    model=model,
    model_kwargs={
        "torch_dtype": torch_dtype,
        "device_map": "auto",
        "attn_implementation": attn_implementation,
    }
)

Run

messages = [
    {"role": "system", "content": "You are a helpful assistant."},
    {"role": "user", "content": "Tell me a recipe for a spicy margarita."},
]
prompt = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
print("***Prompt:\n", prompt)

outputs = pipeline(prompt, max_new_tokens=1000, do_sample=True, temperature=0.7, top_k=50, top_p=0.95)
print("***Generation:\n", outputs[0]["generated_text"][len(prompt):])

Output

"""***Prompt:
 <|im_start|>system
You are a helpful assistant.<|im_end|>
<|im_start|>user
Tell me a recipe for a spicy margarita.<|im_end|>
<|im_start|>assistant

***Generation:
 Sure! Here's a recipe for a spicy margarita:

Ingredients:

- 2 oz silver tequila
- 1 oz triple sec
- 1 oz fresh lime juice
- 1/2 oz simple syrup
- 1/2 oz fresh lemon juice
- 1/2 tsp jalapeño, sliced (adjust to taste)
- Ice cubes
- Salt for rimming the glass

Instructions:

1. Prepare the glass by running a lime wedge around the rim of the glass. Dip the rim into a shallow plate of salt to coat.
2. Combine the tequila, triple sec, lime juice, simple syrup, lemon juice, and jalapeño slices in a cocktail shaker.
3. Add ice cubes to the cocktail shaker and shake vigorously for 30 seconds to 1 minute.
4. Strain the cocktail into the prepared glass.
5. Garnish with a lime wedge and jalapeño slice.

Enjoy! This spicy margarita has a nice balance of sweetness and acidity, with a subtle heat from the jalapeño that builds gradually as you sip."""

🏆 Evaluation

See src/eval for all evaluation runs.

We evaluate models on 6 key benchmarks using Eleuther.AI's Language Model Evaluation Harness.

1. dfurman/Llama-2-70B-Instruct-v0.1

• Precision: fp16

Metric	Value
Avg.	65.72
ARC (25-shot)	69.62
HellaSwag (10-shot)	86.82
MMLU (5-shot)	69.18
TruthfulQA (0-shot)	57.43
Winogrande (5-shot)	83.9
GSM8K (5-shot)	27.37

🤝 References

We finetune off of the following base models:

• llama-3
• phi-2
• mixtral
• mistral
• llama-2
• falcon

We use the following datasets:

• mlabonne/orpo-dpo-mix-40k
• ehartford/dolphin
• jondurbin/airoboros-2.2.1
• garage-bAInd/Open-Platypus
• timdettmers/openassistant-guanaco

We use the following compute providers:

• RunPod
• Lambda Labs
• Google Colab

---