iravikr/distilgpt2-movieqa

DistilGPT2 Movie Question Answering Model

Overview

This repository contains a fine-tuned DistilGPT2 model designed to answer movie-related questions.
The model can respond to queries about movie plots, directors, cast, and general trivia using a simple prompt format:

The goal of this project is to demonstrate domain-specific fine-tuning of a lightweight language model using publicly available movie datasets.

---

Model Details

• Model type: Causal Language Model (Text Generation)
• Base model: distilbert/distilgpt2
• Language: English
• License: Apache 2.0
• Fine-tuning objective: Movie Question Answering

Author

• Name: Ravi Kumar
• Hugging Face: https://huggingface.co/iravikr
• GitHub: https://github.com/dcsgod
• LinkedIn: https://linkedin.com/in/ravi3kr
• Email: rk9128557489@gmail.com

---

Training Data

The model was fine-tuned using the following datasets from Hugging Face:

1. HiTruong/movie_QA

   • Contains curated movie-related question–answer pairs.
   • Used as direct supervised QA training data.

2. Pablinho/movies-dataset

   • Contains movie metadata such as plot summaries and descriptions.
   • Converted into question–answer format during preprocessing.

All training samples were normalized into the following text format:

---

Training Procedure

Preprocessing

• Converted structured metadata into natural language Q&A pairs.
• Tokenized using the DistilGPT2 tokenizer.
• Maximum sequence length: 256 tokens.
• Padding token set to EOS token.

Hyperparameters

• Training regime: fp16 mixed precision
• Optimizer: AdamW
• Learning rate: 5e-5
• Epochs: 3
• Effective batch size: 16 (via gradient accumulation)
• Training steps: 13,095

Compute

• Hardware: NVIDIA T4 GPU
• Platform: Google Colab
• Training time: ~1.1 hours

Training Result

• Final training loss: ~2.29
• Loss steadily decreased, indicating stable learning without collapse or overfitting.

---

Intended Use

Direct Use

• Movie question answering chatbots
• Movie trivia applications
• Educational demos
• Lightweight domain-specific assistants

Downstream Use

• Can be combined with retrieval systems (FAISS / Chroma) for RAG-based movie QA
• Can be deployed via FastAPI or Streamlit
• Suitable for experimentation and learning purposes

Out-of-Scope Use

• Real-time or up-to-date movie facts without retrieval
• Complex multi-hop reasoning
• High-stakes or authoritative factual systems
• Long conversational memory use cases

---

Limitations and Risks

• The model may hallucinate facts for movies not present in the training data.
• Knowledge is limited to the datasets used during fine-tuning.
• Bias may exist toward popular or English-language films.

Recommendations

• Use retrieval augmentation for production systems.
• Validate responses when factual correctness is critical.
• Avoid using the model as a sole source of truth.

---

How to Use

Loading the Model

from transformers import pipeline

pipe = pipeline(
    "text-generation",
    model="iravikr/distilgpt2-movieqa",
    tokenizer="iravikr/distilgpt2-movieqa"
)

pipe(
    "Question: Who directed Inception?\nAnswer:",
    max_new_tokens=40,
    temperature=0.7,
    top_p=0.9
)
Question: Who directed Inception?
Answer: Inception was directed by Christopher Nolan.

---

## Training Procedure

### Preprocessing

- Converted structured movie metadata into natural language QA pairs.
- Tokenized using the DistilGPT2 tokenizer.
- Maximum sequence length: 256 tokens.
- Padding token set to EOS token.

### Hyperparameters

- **Training regime:** fp16 mixed precision
- **Optimizer:** AdamW
- **Learning rate:** 5e-5
- **Epochs:** 3
- **Effective batch size:** 16 (via gradient accumulation)
- **Total training steps:** 13,095

---

## Hardware and Compute

- **GPU:** NVIDIA T4 (16 GB VRAM)
- **CPU:** Intel Xeon (Google Colab backend)
- **RAM:** ~12 GB
- **Platform:** Google Colab
- **Training time:** Approximately 1.1 hours

---

## Training Outcome

- **Initial training loss:** ~2.98  
- **Final training loss:** ~2.29  

The steady loss reduction indicates stable convergence without overfitting.

---

## Intended Use

### Direct Use

- Movie question answering systems
- Movie trivia bots
- Educational and demonstration projects
- Lightweight domain-specific assistants

### Downstream Use

- Integration with retrieval systems (FAISS / Chroma) for RAG-based QA
- Deployment via FastAPI or Streamlit
- Extension for recommendation-style prompts

### Out-of-Scope Use

- Real-time or up-to-date movie facts without retrieval
- Complex multi-hop reasoning
- High-stakes decision-making systems
- Long-context conversational memory

---

## Limitations and Risks

- The model may hallucinate facts for movies not seen during training.
- Knowledge is limited to the scope of the training datasets.
- Dataset bias toward popular and English-language films may exist.

### Recommendations

- Use retrieval augmentation for production deployments.
- Validate responses for factual accuracy.
- Avoid use as a sole authoritative source.

---

## How to Use

### Loading the Model

```python
from transformers import pipeline

pipe = pipeline(
    "text-generation",
    model="iravikr/distilgpt2-movieqa",
    tokenizer="iravikr/distilgpt2-movieqa"
)

pipe(
    "Question: Who directed Inception?\nAnswer:",
    max_new_tokens=40,
    temperature=0.7,
    top_p=0.9
)