cwestnedge/gpt2-large-pubmed

Overview

This pipeline was used to fine‑tune GPT‑2 small, medium, and large on abstracts from PubMed's baseline data. Models were trained on a single A100 GPU in Google Colab.

---

Training

Setup

• Single epoch over 221,709 batches × 16 × 1024 tokens ≈ 3.63 billion tokens
• Identical optimizer, learning‑rate schedule, and hyper‑parameters for all models
• No additional regularization or early stopping

Loss

Here are the the loss curves for GPT‑2 small, medium, and large fine‑tuned on PubMed abstracts over single epoch.

• Loss comparisons

---

Evaluation

Dataset

Hold‑out set of 1000 × 16 × 1024 tokens (≈ 16.4 M tokens) randomly sampled from PubMed abstracts, disjoint from the training split.

Metrics

Cross‑entropy loss (averaged over all tokens) and derived perplexity (ppl = exp(loss)) on the hold‑out set:

Model	Parameters	Avg CE Loss ↓	Perplexity ↓
gpt2‑small‑pubmed	124 M	2.5017	12.20
gpt2‑medium‑pubmed	355 M	2.2984	9.96
gpt2‑large‑pubmed	774 M	2.1863	8.90

Caveats

• Perplexities are in‑domain (PubMed abstracts) and may not reflect general‑purpose LM quality
• Only one epoch of training; performance likely improves with more epochs or hyper‑parameter tuning
• Downstream biomedical benchmarks have not yet been conducted

---

Usage

1) Quick‑start with the 🤗 pipeline API

from transformers import pipeline
import torch 

device = "cuda" if torch.cuda.is_available() else "cpu"

generator = pipeline(
    "text-generation",
    model="cwestnedge/gpt2-large-pubmed",
    tokenizer="openai-community/gpt2-large",
    device=device,
)

prompt = (
    "Background: The CRISPR–Cas9 system has revolutionized gene editing. "
    "In this study, we evaluate its efficacy in"
)
out = generator(
    prompt,
    max_length=200,
    temperature=1e-9,
    top_p=1e-9,
    num_return_sequences=1,
    truncation=True,
)
print(out[0]["generated_text"])

2) Manual load + generate for finer control

import torch
from transformers import AutoTokenizer, AutoModelForCausalLM

device = "cuda" if torch.cuda.is_available() else "cpu"
model_name = "cwestnedge/gpt2-large-pubmed"
tokenizer = AutoTokenizer.from_pretrained("openai-community/gpt2-large")
model = AutoModelForCausalLM.from_pretrained(model_name).to(device)

inputs = tokenizer(
    "Methods: We performed a double‐blind randomized trial to assess",
    return_tensors="pt",
).to(device)

gen_ids = model.generate(
    **inputs,
    max_length=150,
    num_beams=5,
    no_repeat_ngram_size=2,
    early_stopping=True,
)
print(tokenizer.decode(gen_ids[0], skip_special_tokens=True))

3) Scoring / perplexity

import torch
from transformers import AutoTokenizer, AutoModelForCausalLM

device = "cuda" if torch.cuda.is_available() else "cpu"
model_name = "cwestnedge/gpt2-large-pubmed"
tokenizer = AutoTokenizer.from_pretrained("openai-community/gpt2-large")
model = AutoModelForCausalLM.from_pretrained(model_name).to(device)

text = (
    "Tetralogy of Fallot is a rare congenital heart condition that is present at birth."
)
enc = tokenizer(text, return_tensors="pt").to(device)

with torch.no_grad():
    outputs = model(**enc, labels=enc.input_ids)
    loss = outputs.loss
    ppl = torch.exp(loss)

print(f"CE loss: {loss:.4f} → Perplexity: {ppl:.2f}")