Nihal108-bi/tinyllama-math-instruct-lora

library_name: transformers tags: - peft - lora - tinyllama - instruction-tuning - mathematics - machine-learning - education - domain-adaptation - two-stage-training - gsm8k license: apache-2.0

🧠 TinyLlama Math & ML Tutor (LoRA)

A two-stage fine-tuned educational model specialized in Mathematics and Machine Learning

Built on TinyLlama-1.1B — lightweight, fast, and effective for learning & teaching.

---

🚀 Training Strategy (Two-Stage Fine-Tuning)

This model was developed using a two-stage fine-tuning approach for maximum mathematical depth and instruction-following ability.

🥇 Stage 1: Domain Adaptation (Non-Instruction)

• Dataset: Full textbook — Mathematics of Machine Learning: Master Linear Algebra, Calculus, and Probability for Machine Learning by Tivadar Danka
• Method: Continued pre-training (non-instruction / causal language modeling)
• Goal: Inject deep mathematical knowledge, terminology, and formal reasoning into the base model.

🥈 Stage 2: Instruction Tuning

• Dataset: openai/gsm8k (main split)
• Method: Instruction tuning using [INST] format
• Goal: Teach the model to answer user questions clearly, step-by-step, and in an educational style.

Result: Strong domain expertise + natural instruction following.

---

🧾 Model Details

• Base Model: TinyLlama/TinyLlama-1.1B-intermediate-step-1431k-3T
• Fine-tuning Method: LoRA (PEFT)
• LoRA Rank (r): 8
• Domain: Mathematics + Machine Learning
• Language: English
• Developed by: Nihal Jaiswal

---

🧠 Prompt Format

Use the following format for best results:

[INST]Your question here[/INST]

---

💬 Example Usage

Input:

[INST]What is a vector space? Explain with example.[/INST]

Output: A vector space is a mathematical structure consisting of a set of vectors that can be added together and multiplied by scalars...

---

⚙️ How to Use (Copy-Paste)

from transformers import AutoModelForCausalLM, AutoTokenizer
from peft import PeftModel

# Load model
repo_id = "Nihal108-bi/tinyllama-math-instruct-lora"
base_model = "TinyLlama/TinyLlama-1.1B-intermediate-step-1431k-3T"

tokenizer = AutoTokenizer.from_pretrained(repo_id)
model = AutoModelForCausalLM.from_pretrained(base_model, device_map="auto", torch_dtype="auto")
model = PeftModel.from_pretrained(model, repo_id)

# Generate
prompt = "[INST]Explain gradient descent with an example[/INST]"
inputs = tokenizer(prompt, return_tensors="pt").to("cuda")

outputs = model.generate(
    **inputs,
    max_new_tokens=300,
    temperature=0.7,
    top_p=0.9,
    do_sample=True,
    repetition_penalty=1.1
)

answer = tokenizer.decode(outputs[0], skip_special_tokens=True).split("[/INST]")[-1].strip()
print(answer)

---

🎯 Capabilities

• Strong explanations of Linear Algebra, Calculus, Probability & Statistics
• Clear Machine Learning concepts (Gradient Descent, Bias-Variance, etc.)
• Step-by-step mathematical reasoning
• Educational / tutor-style responses

⚠️ Limitations

• Small 1.1B model → can occasionally hallucinate on very advanced topics
• Best performance on math and ML-related questions
• Not designed for general chit-chat or coding

---

📊 Evaluation

Qualitatively evaluated on math & ML reasoning tasks.

Category	Performance
Math Concepts	Strong
ML Explanations	Strong
Step-by-step Reasoning	Good
General Conversation	Limited

---

🙌 Credits & Acknowledgments

• Fine-tuned by: Nihal Jaiswal
• Base Model: TinyLlama Team
• Domain Dataset: Tivadar Danka – Mathematics of Machine Learning
• Instruction Dataset: OpenAI GSM8K

---

📜 License

Apache 2.0

---

Made with ❤️ for students, educators, and ML enthusiasts.

Feel free to use, share, and build upon it!