phi1-numpy-pandas-qlora

Phi-1 NumPy + Pandas Fine-Tuning (QLoRA)

This repository provides a fine-tuned variant of :contentReference[oaicite:0]{index=0} that improves Python code generation for NumPy and Pandas–based data analysis tasks.

The model addresses a known limitation of the base Phi-1 model, which was trained primarily on Python scripts using a very limited set of standard libraries.

---

🔍 Motivation

The original Phi-1 model documentation states that 99.8% of its Python training data only uses:

As a result, the base model may:

• Hallucinate APIs for external libraries
• Misuse common NumPy / Pandas functions
• Struggle with real-world data analysis workflows

This project demonstrates that targeted, low-rank fine-tuning (QLoRA) can significantly mitigate this limitation.

---

🧠 What This Model Learns

This fine-tuning focuses exclusively on:

• numpy
• pandas

Covered Operations

• CSV loading (read_csv)
• Column selection and filtering
• GroupBy aggregations (mean, sum, count)
• NumPy array operations
• Pandas + NumPy interoperability (np.where, .to_numpy())

Explicitly NOT Included

• Visualization libraries (matplotlib, seaborn)
• Machine learning libraries (sklearn, torch)
• SQL, file systems, or OS operations

---

🏗️ Training Details

• Base model: microsoft/phi-1
• Fine-tuning method: QLoRA (4-bit quantization)
• LoRA rank (r): 16
• Epochs: 2–3
• Dataset size: ~1,500 instruction–code pairs
• Training type: Instruction tuning (code-only responses)
• Optimizer: paged AdamW (8-bit)

Only LoRA adapter weights were trained; the base model weights remain unchanged.

---

📦 Repository Contents

Depending on the repository variant, this repo contains:

Adapter Repository (Recommended)

• adapter_model.bin (or .safetensors)
• adapter_config.json
• Tokenizer files

Full Model Repository (If merged)

• Full merged model weights
• Tokenizer
• No PEFT dependency required

---

Usage

Option 1 — Load Adapter (Recommended)

from transformers import AutoModelForCausalLM, AutoTokenizer
from peft import PeftModel

base_model = AutoModelForCausalLM.from_pretrained(
    "microsoft/phi-1",
    load_in_4bit=True,
    device_map="auto"
)

model = PeftModel.from_pretrained(
    base_model,
    "iravikr/phi1-numpy-pandas-qlora"
)

tokenizer = AutoTokenizer.from_pretrained(
    "iravikr/phi1-numpy-pandas-qlora"
)