Gokul Biodata FLAN-T5
Model Description
A small demo model: google/flan-t5-base fully fine-tuned to answer questions about the resume of Gokul S, an MSc Data Science student at Dayananda Sagar University, Bengaluru.
It is a personal-data demonstration of memorizing a handful of resume facts (name, summary, education, school, programming languages, etc.). It is not a general-purpose model, and it will likely give wrong or made-up answers for anything outside those facts.
- Developed by: Gokul S
- Model type: Encoder-decoder (seq2seq) transformer
- Language: English
- License: Apache 2.0 (same as the base model)
- Finetuned from: google/flan-t5-base
Training Details
- Method: Full fine-tuning (all weights trainable)
- Data: 30 hand-written question-answer pairs built from resume facts (3 question phrasings per fact), plus 10 held-out rephrased questions used as a check set
- Epochs: 40
- Learning rate: 3e-4
- Batch size: 4
- Hardware: Google Colab T5 GPU
- Framework: Hugging Face Transformers (
Seq2SeqTrainer)
How to Use
The model expects this exact prompt format:
from transformers import AutoTokenizer, AutoModelForSeq2SeqLM
repo = "gokul2124/gokul-biodata-flan-t5"
tok = AutoTokenizer.from_pretrained(repo)
model = AutoModelForSeq2SeqLM.from_pretrained(repo)
PROMPT = "Answer the question about Gokul S using his resume. Question: {} Answer:"
question = "What is your name?"
inputs = tok(PROMPT.format(question), return_tensors="pt", truncation=True, max_length=128)
out = model.generate(**inputs, max_new_tokens=128, do_sample=False)
print(tok.decode(out[0], skip_special_tokens=True))
Limitations
- Trained on a tiny dataset, so it only reliably reproduces the facts it was trained on.
- Can hallucinate or give incomplete answers for unseen questions.
- Low loss means it fits the supplied answers, not that it is factually accurate in general.
- Contains personal information from the author's resume. Contact details (phone, email, address) were deliberately excluded.
Intended Use
Learning and demonstration of fine-tuning a seq2seq model on custom Q&A data. Not intended for any production or decision-making use.