This repository contains a LoRA adapter (not a full merged model) for GPT-2, fine-tuned on the WikiText-2 raw v1 dataset for causal language modelling using the PEFT library.

Important: This repo contains only the LoRA adapter delta weights. Inference requires loading the GPT-2 base model separately and applying the adapter on top (see usage instructions below).

Model Details

Field	Value
Base model	`gpt2`
Adapter type	LoRA (Low-Rank Adaptation)
Dataset	WikiText-2 raw v1
Task	Causal Language Modelling
Frameworks	Transformers, PEFT, PyTorch
Training environment	Google Colab — Tesla T4 GPU
Selected checkpoint	checkpoint-1752

Evaluation Results

Metric	Value
Best validation loss	3.1842
Best validation perplexity	24.1475
Test loss	3.1593
Test perplexity	23.5548
Target baseline (val PPL)	18 – 25
Baseline achieved	✓ Yes

Usage

python

from transformers import AutoModelForCausalLM, AutoTokenizer
from peft import PeftModel

# 1. Load the GPT-2 base model
base_model = AutoModelForCausalLM.from_pretrained("gpt2")

# 2. Load the tokenizer from this repo
tokenizer = AutoTokenizer.from_pretrained("dina1/gpt2-wikitext2-lora-peft",
                                          subfolder="tokenizer")

# 3. Apply the LoRA adapter
model = PeftModel.from_pretrained(base_model, "dina1/gpt2-wikitext2-lora-peft")
model.eval()

# 4. Generate text
inputs = tokenizer("The history of artificial intelligence", return_tensors="pt")
outputs = model.generate(**inputs, max_new_tokens=100, do_sample=True, temperature=0.9)
print(tokenizer.decode(outputs[0], skip_special_tokens=True))

Reproducibility

Hyperparameter	Value
Block size	512 tokens
LoRA rank (r)	8
LoRA alpha	16
LoRA dropout	0.1
LoRA target modules	`["c_attn"]`
Learning rate	2e-4
LR scheduler	cosine
Gradient accumulation	8 steps
Training precision	fp16
Seed	42

Limitations

Trained exclusively on WikiText-2 (English Wikipedia/news articles).
Not instruction-tuned or chat-aligned.
Not safety-aligned — outputs may be factually incorrect or biased.
Intended for educational and experimental use only.
Performance on domains outside WikiText-2 is not guaranteed.

Files in This Repository

File / Directory	Description
`adapter_config.json`	LoRA adapter configuration
`adapter_model.safetensors`	LoRA adapter weights
`tokenizer/`	GPT-2 tokenizer files
`evaluation_summary.json`	Full evaluation metrics from Phase 5
`production_manifest.json`	Deployment metadata and artifact paths
`loss_and_perplexity_curves.png`	Training and validation curves

Citation

If you use this adapter in your work, please credit the PEFT library:

bibtex

@misc{peft,
  author = {Sourab Mangrulkar and Sylvain Gugger and Lysandre Debut and
            Younes Belkada and Sayak Paul},
  title  = {PEFT: State-of-the-art Parameter-Efficient Fine-Tuning methods},
  year   = {2022},
  url    = {https://github.com/huggingface/peft}
}