Harsha901/Qwen3-4B-Inst-Math-Reasoning-SFT

Uploaded Model

• Developed by: Harsha901
• License: apache-2.0
• Finetuned from model: unsloth/Qwen3-4B-Instruct-2507

This Qwen3 model was trained ~2× faster using Unsloth and Hugging Face’s TRL library.

---

📌 Model Overview

Qwen3-4B-Inst-Math-Reasoning-SFT is a supervised fine-tuned (SFT) variant of Qwen3-4B-Instruct, optimized for mathematical reasoning and step-by-step problem solving.

The model is trained to follow instructions precisely while producing clear, logically structured reasoning chains, making it suitable for:

• Math problem solving
• Educational assistants
• Reasoning benchmarks
• Downstream alignment (DPO / RLHF)

---

🧠 Key Capabilities

• Multi-step mathematical reasoning
• Algebra, arithmetic, and word problems
• Chain-of-thought style explanations
• Improved instruction adherence
• More stable reasoning compared to the base model

---

🏗️ Model Architecture

• Architecture: Decoder-only Transformer (Causal LM)
• Parameters: ~4B
• Base Model: Qwen3-4B-Instruct (Unsloth optimized)
• Tokenization: Qwen tokenizer
• Context Length: Same as base model

---

📚 Training Data

The model was fine-tuned on a curated dataset consisting of:

• Instruction-style math prompts
• Step-by-step mathematical solutions
• Reasoning-focused explanations

Data was filtered to emphasize:

• Logical consistency
• Clear intermediate steps
• Reduced ambiguity in solutions

While care was taken to ensure quality, the dataset may still contain noise or biases present in public mathematical corpora.

---

⚙️ Training Details

• Fine-tuning Method: Supervised Fine-Tuning (SFT)
• Frameworks: Hugging Face Transformers + TRL
• Acceleration: Unsloth (memory-efficient & faster training)
• Precision: FP16 / BF16 (hardware dependent)
• Optimizer: AdamW
• Loss Function: Cross-entropy
• Batching: Gradient accumulation for memory efficiency

---

🚀 Usage

Load the Model

from transformers import AutoModelForCausalLM, AutoTokenizer

model_id = "Harsha901/Qwen3-4B-Inst-Math-Reasoning-SFT"

tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(
    model_id,
    device_map="auto",
    torch_dtype="auto"
)

Example Inference

prompt = "Solve step by step: If 5x − 10 = 15, find x."

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

outputs = model.generate(
    **inputs,
    max_new_tokens=256,
    temperature=0.2,
    do_sample=False
)

print(tokenizer.decode(outputs[0], skip_special_tokens=True))

---

📊 Evaluation

The model was evaluated qualitatively on:

• Math word problems
• Algebraic equations
• Multi-step reasoning tasks

Observed improvements vs base model:

• Better structured reasoning
• More consistent intermediate steps
• Fewer incomplete solutions

Formal benchmark results (e.g., GSM8K, MATH) are planned for future updates.

---

⚠️ Limitations

• Not guaranteed to be mathematically correct in all cases
• Can be verbose due to reasoning-style outputs
• Not optimized for creative or non-technical writing
• Performance may degrade on extremely long or ambiguous prompts

---

🔐 Ethical & Responsible Use

• Intended for research and educational purposes
• Outputs should be verified for correctness in critical applications
• Not suitable for high-stakes decision-making without human oversight

---

📜 License

Released under the Apache 2.0 License, consistent with the base Qwen3 model.

---

🙌 Acknowledgements

• Qwen Team for the base Qwen3 architecture
• Unsloth for efficient fine-tuning optimizations
• Hugging Face for Transformers and TRL

---

✉️ Author

Harsha Vardhan Mannem AI / ML Engineer Hugging Face & GitHub: Harsha901

---

🔮 Future Work

• Preference tuning with DPO
• Quantized inference (4-bit / 8-bit)
• Benchmark-based evaluation
• Deployment-optimized variants