Minueza-32M-Chat: A chat model with 32 million parameters

Base model: Felladrin/Minueza-32M-Base
Datasets used during SFT:
Datasets used during DPO:
License: Apache License 2.0
Availability in other ML formats:
- GGUF: Felladrin/gguf-Minueza-32M-Chat
- ONNX: Felladrin/onnx-Minueza-32M-Chat

Recommended Prompt Format

<|im_start|>system
{system_message}<|im_end|>
<|im_start|>user
{user_message}<|im_end|>
<|im_start|>assistant

Recommended Inference Parameters

do_sample: true
temperature: 0.65
top_p: 0.55
top_k: 35
repetition_penalty: 1.176

Usage Example

from transformers import pipeline

generate = pipeline("text-generation", "Felladrin/Minueza-32M-Chat")

messages = [
    {
        "role": "system",
        "content": "You are a helpful assistant who answers the user's questions with details and curiosity.",
    },
    {
        "role": "user",
        "content": "What are some potential applications for quantum computing?",
    },
]

prompt = generate.tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)

output = generate(
    prompt,
    max_new_tokens=256,
    do_sample=True,
    temperature=0.65,
    top_k=35,
    top_p=0.55,
    repetition_penalty=1.176,
)

print(output[0]["generated_text"])

How it was trained

This model was trained with SFT Trainer and DPO Trainer, in several sessions, using the following settings:

For Supervised Fine-Tuning:

Hyperparameter	Value
learning_rate	2e-5
total_train_batch_size	24
max_seq_length	2048
weight_decay	0
warmup_ratio	0.02

For Direct Preference Optimization:

Hyperparameter	Value
learning_rate	7.5e-7
total_train_batch_size	6
max_length	2048
max_prompt_length	1536
max_steps	200
weight_decay	0
warmup_ratio	0.02
beta	0.1

Open LLM Leaderboard Evaluation Results

Detailed results can be found here

Metric	Value
Avg.	28.49
AI2 Reasoning Challenge (25-Shot)	20.39
HellaSwag (10-Shot)	26.54
MMLU (5-Shot)	25.75
TruthfulQA (0-shot)	47.27
Winogrande (5-shot)	50.99
GSM8k (5-shot)	0.00