🧠 Rust-Master-thinking

This repository contains a fine-tuned version of unsloth/phi-4-reasoning, trained with LoRA on the Tesslate/Rust_Dataset. The goal of this project is to enhance the model's reasoning, explanation, and step-by-step thinking abilities specifically for Rust-related tasks.

🚀 Model Purpose

This model was fine-tuned to:

• Improve Rust coding explanations
• Generate high-quality reasoning traces
• Provide step-by-step problem solving
• Give detailed and structured answers

The training format follows:

<|user|>
{prompt}
<|assistant|>
<think>
{reasoning}
</think>
{response}

🔧 How to Use

Install dependencies (if not installed):

pip install transformers bitsandbytes

Load model normally:

from transformers import AutoTokenizer, AutoModelForCausalLM
import torch

model_id = "SkyAsl/Rust-Master-thinking"

tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(model_id, dtype=torch.bfloat16, device_map="auto")
model.eval()

prompt = "Explain why Rust ownership prevents data races."

input_text = (
    f"<|user|>\n{prompt}\n"
    f"<|assistant|>\n<think>\n"
)

inputs = tokenizer(input_text, return_tensors="pt").to(model.device)

with torch.no_grad():
    output = model.generate(
        **inputs,
        max_new_tokens=3000,
        temperature=0.7,
        top_p=0.9,
        do_sample=True,
        repetition_penalty=1.2,
    )

print(tokenizer.decode(output[0], skip_special_tokens=False))

🧩 Base Model

unsloth/phi-4-reasoning

• 14B parameter reasoning-optimized model
• Uses internal <think> reasoning
• Strong on step-by-step chain-of-thought tasks

🛠 Fine-Tuning Details

Setting	Value
Method	LoRA (PEFT)
Rank (r)	16
Alpha	32
Dropout	0.05
Target Modules	q/k/v/o proj, mlp (up/down/gate)
Max Length	512
Precision	4-bit QLoRA
Batch Size	16
Grad Accum	8
LR	2e-4
Scheduler	cosine
Epochs	1

🤖 Evaluation

Epoch	Training Loss	Validation Loss
1	2.251500	2.191743

📚 Dataset

Tesslate/Rust_Dataset

Includes:

• Rust prompts
• Step-by-step reasoning
• Final answers

This dataset improves the model's ability to produce structured and accurate explanations for Rust programming tasks.