Siddharth466/tinyllama-gsm8k-math-lora

TinyLlama GSM8K Math Fine-Tuned (QLoRA)

This model is a LoRA fine-tuned adapter built on top of:

Base Model: TinyLlama/TinyLlama-1.1B-Chat-v1.0
Dataset: GSM8K (Grade School Math 8K)
Fine-tuning Method: QLoRA (4-bit quantization + LoRA)

It improves step-by-step mathematical reasoning and structured problem-solving.

---

Training Details

• Base Model: TinyLlama-1.1B-Chat
• Method: QLoRA (4-bit NF4 quantization)
• LoRA Rank (r): 16
• LoRA Alpha: 32
• LoRA Dropout: 0.05
• Epochs: 3
• Learning Rate: 2e-4
• Optimizer: paged_adamw_8bit
• Scheduler: Cosine
• Max Sequence Length: 512
• Dataset: GSM8K (train split)

Comparision

Evaluating Base TinyLlama...

100%|██████████| 165/165 [12:35<00:00, 4.58s/it] Base TinyLlama Accuracy: 1.29%

Evaluating Fine-Tuned (Merged)...

100%|██████████| 165/165 [12:40<00:00, 4.61s/it] Fine-Tuned (Merged) Accuracy: 1.90%

Training was performed using:

• Transformers
• PEFT
• TRL SFTTrainer
• BitsAndBytes

---