rohanjain11/flan-t5-mlds-qlora

Flan-T5 ML/Data Science QLoRA Adapter

A QLoRA (LoRA) adapter fine-tuned on top of `google/flan-t5-base` for machine learning and data science question answering. This repo contains only the adapter weights (~16 MB), not the full base model.

Model Description

This model adapts Flan-T5-base to answer practitioner-style ML/Data Science questions with longer, domain-relevant responses. Training used a synthetic Q&A dataset (324 examples) generated with OpenAI gpt-4o-mini, then QLoRA fine-tuning on Google Colab (T4 GPU).

• —Developed by: rohanjain11
• —Model type: Seq2seq language model (LoRA adapter)
• —Language: English
• —License: Apache 2.0 (inherits from Flan-T5-base)
• —Finetuned from: `google/flan-t5-base`
• —Training method: QLoRA (8-bit base + LoRA on attention projections)
• —Code / dataset pipeline: github.com/rohanjain11/llm-finetune-qlora

Intended Use

Direct Use

Answering English questions about machine learning and data science topics (model evaluation, cross-validation, neural networks, feature engineering, deployment, etc.) in an instructional Q&A format.

Example prompt format used during training:

Answer the following question about machine learning and data science: {question}

Downstream Use

• —Domain-specific QA assistant (with human review)
• —Starting point for further LoRA fine-tuning on related technical domains
• —Educational / portfolio demonstration of parameter-efficient fine-tuning

Out-of-Scope Use

• —Medical, legal, or financial advice
• —High-stakes decisions without human verification
• —Factual claims requiring up-to-date sources (model may hallucinate)
• —Languages other than English

Bias, Risks, and Limitations

• —Trained on synthetic data from an LLM; answers may reflect generator biases or inaccuracies.
• —Small fine-tuning set (324 examples); coverage is broad but not exhaustive.
• —Flan-T5-base capacity limits depth vs. much larger models.
• —ROUGE-L improved on a held-out validation set but does not guarantee factual correctness.
• —Not evaluated for fairness across demographic groups or sensitive topics.

Recommendation: Treat outputs as draft educational content; verify important facts against primary sources.

How to Get Started

import torch
from peft import PeftModel
from transformers import AutoModelForSeq2SeqLM, AutoTokenizer

BASE_MODEL = "google/flan-t5-base"
ADAPTER = "rohanjain11/flan-t5-mlds-qlora"

tokenizer = AutoTokenizer.from_pretrained(ADAPTER)
base_model = AutoModelForSeq2SeqLM.from_pretrained(
    BASE_MODEL,
    torch_dtype=torch.float16,
    device_map="auto",
)
model = PeftModel.from_pretrained(base_model, ADAPTER)
model.eval()

question = "What is the bias-variance tradeoff?"
prompt = f"Answer the following question about machine learning and data science: {question}"

inputs = tokenizer(prompt, return_tensors="pt").to(model.device)
outputs = model.generate(
    **inputs,
    max_new_tokens=256,
    min_length=40,
    num_beams=4,
    early_stopping=True,
)
print(tokenizer.decode(outputs[0], skip_special_tokens=True))

Training Details

Training Data

• —Source: Synthetic Q&A pairs generated locally via OpenAI API (gpt-4o-mini)
• —Domain: Machine learning & data science (40 subtopics)
• —Train examples: 324
• —Validation examples: 36
• —Format: JSONL with instruction (question) and output (answer) fields
• —Generation cost: ~$0.04 (dataset creation)

Training Procedure

Preprocessing

• —Input: Answer the following question about machine learning and data science: {instruction}
• —Target: {output}
• —Max input length: 256 tokens
• —Max target length: 400 tokens
• —Label padding masked with -100

Training hyperparameters

Speeds, sizes, times

Loss (teacher-forced)

Evaluation

Testing Data

36 held-out validation examples from the same synthetic dataset (10% split, random_state=42).

Metrics

ROUGE-L (longest common subsequence F-measure) comparing generated answers to reference answers on the full validation set, using min_length=40 and beam search (num_beams=4).

Base model evaluated as a separate google/flan-t5-base load (not adapter-disabled PEFT).

Results

Qualitative outputs improved in length and domain tone vs. base, but may still contain inaccuracies typical of small models on open-ended QA.

Environmental Impact

Approximate Colab T4 usage for fine-tuning only:

• —Hardware: NVIDIA T4
• —Training time: ~0.04 hours (~2.2 minutes)
• —Cloud provider: Google Colab (free tier)
• —Carbon: Negligible for a single short fine-tuning run

Technical Specifications

Model Architecture

• —Architecture: T5 encoder-decoder (Flan-T5-base)
• —Objective: Conditional generation (seq2seq LM)
• —This repo: PEFT LoRA adapter weights only

Compute Infrastructure

• —Hardware: Google Colab T4 GPU
• —Software: Python 3.12, PyTorch, Transformers, PEFT, bitsandbytes, Accelerate

Citation

Base model — Flan-T5:

@misc{chung2022scaling,
  title={Scaling Instruction-Finetuned Language Models},
  author={Hyung Won Chung and others},
  year={2022},
  eprint={2210.11416},
  archivePrefix={arXiv},
  primaryClass={cs.LG}
}

LoRA:

@inproceedings{hu2022lora,
  title={LoRA: Low-Rank Adaptation of Large Language Models},
  author={Edward J. Hu and others},
  booktitle={ICLR},
  year={2022}
}

Model Card Contact

rohanjain11 · GitHub