DataSci-Coder-14B-LoRA

DataSci-Coder-14B: Qwen2.5-Coder-14B LoRA Adapter for Data Science

A QLoRA fine-tuned adapter for Qwen2.5-Coder-14B-Instruct optimized for data science code generation. The model outputs clean, runnable Python code with zero explanatory text — strictly following code-only instructions.

Key Results

What It Does

• Generates complete, runnable Python code for data science tasks
• Covers statistics, machine learning, deep learning, NLP, time series, and visualization
• Follows instructions precisely — when told "no explanations," it outputs only code (base model ignores this 40% of the time)
• Handles complex tasks: Bayesian inference, VAEs, GANs, survival analysis, stacking ensembles, SHAP, anomaly detection

Training Details

Training Data

10,795 curated data science instruction-response pairs from:

• 6 public HuggingFace datasets (CodeAlpaca, Evol-Instruct, etc.)
• University coursework (statistics, ML, deep learning)
• Data science newsletters
• Hand-curated examples

All examples filtered for Python code quality, data science relevance, and length. Categories: machine learning, deep learning, statistics, data wrangling, visualization, NLP, time series, numerical computing.

Usage

from unsloth import FastLanguageModel
import torch

model, tokenizer = FastLanguageModel.from_pretrained(
    model_name="jsmall12/DataSci-Coder-14B-LoRA",
    max_seq_length=2048,
    load_in_4bit=True,
    dtype=None,
)
FastLanguageModel.for_inference(model)

messages = [
    {"role": "system", "content": "You are an expert data science coding assistant. Respond ONLY with clean, runnable Python code. Use inline comments for explanation. No text outside code blocks."},
    {"role": "user", "content": "Write a function to train a logistic regression model with sklearn and print the classification report."},
]

text = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
inputs = tokenizer(text, return_tensors="pt").to("cuda")

with torch.no_grad():
    output = model.generate(
        **inputs,
        max_new_tokens=1024,
        temperature=0.1,
        do_sample=True,
        top_p=0.9,
        repetition_penalty=1.15,
        use_cache=False,
    )

response = tokenizer.decode(output[0][inputs["input_ids"].shape[1]:], skip_special_tokens=True)
print(response)

Evaluation

Hard Eval (12 Complex Tasks)

All 12 tasks produced correct, complete, runnable implementations:

Constraint Eval (10 Multi-Constraint Tests)

Hardware Requirements

• Minimum: ~10GB VRAM (4-bit quantized)
• Recommended: 24GB+ VRAM (L4, A100, etc.)
• Tested on: NVIDIA L40S (44GB), NVIDIA T4 x2 (15GB each)

License

Apache 2.0