---
license: cc-by-nc-3.0
base_model: unsloth/phi-4-mini-reasoning
tags:
  - text-generation-inference
  - transformers
  - unsloth
  - phi-4
language:
  - en
---

# Phi-4 Mini Reasoning – JEE Mathematics Finetuned Model

A new version of the model present at `harsh762011/numiano14`.

# Uploaded Finetuned Model

- **Developed by:** Harsh Srivastava
- **License:** cc-by-nc-3.0
- **Finetuned from model:** unsloth/phi-4-mini-reasoning

This Phi-4 model was trained faster using Unsloth and Hugging Face TRL.

[<img src="https://raw.githubusercontent.com/unslothai/unsloth/main/images/unsloth%20made%20with%20love.png" width="200"/>](https://github.com/unslothai/unsloth)

---

# Description

This model is a finetuned version of Phi-4 Mini Reasoning designed for solving JEE-level mathematics problems.

The model is optimized for:
- Step-by-step mathematical reasoning
- Symbolic problem solving
- Competitive exam-style question solving

---

# Training Dataset

Total samples used: 500k+ filtered mathematics and reasoning samples.

The training pipeline focuses on JEE-level mathematical difficulty using keyword-based dataset filtering.

## Sources

- AI-MO/NuminaMath-CoT — 293k samples (2 epochs)
- AI-MO/NuminaMath-TIR — 68,850 samples
- MetaMathQA — 70k samples
- TIGER-Lab MathInstruct — 125,220 samples
- PhysicsWallahAI JEE Main 2025 (Jan) — 182 samples
- PhysicsWallahAI JEE Main 2025 (Apr) — 169 samples
- MMLU High School Mathematics — 78 samples
- MMLU College Mathematics — 50 samples
- MMLU Abstract Algebra — 25 samples

---

# Training Details

- **Base model:** Phi-4 Mini Reasoning
- **Framework:** Unsloth + Hugging Face TRL
- **Training method:** LoRA finetuning
- **Sequence length:** 2048
- **Optimizer:** AdamW 8bit

---

# Intended Purpose

The model is designed for:

- JEE mathematics reasoning
- Step-by-step mathematical explanations
- Competitive exam problem solving
- Mathematical chain-of-thought reasoning

---

# Limitations

- The model may still generate incorrect mathematical reasoning.
- Outputs should be verified for high-stakes usage.
- The model is still under active improvement and continued training.