Model	Emotion Accuracy	Improvement vs Baseline
Prefix-Small (This Model)	38.2% ⭐	+9.8pp
Token-Small	30.8%	+2.5pp
Prefix-Medium (355M)	35.6%	+7.3pp
Baseline (no conditioning)	28.3%	-
Random Baseline	14.3%	-

Key Achievements:

✅ 2.67x better than random chance
✅ 34.6% relative improvement over baseline
✅ More efficient than larger models (Small > Medium with limited data)
✅ Competitive with literature (35-48% range for similar tasks)

Generation Quality

Perplexity: 24.5 (maintained fluency)
Emotion Success Rate: 38.2% match target emotion
Method: Natural language prefixes (no architecture changes)

Usage

python

from transformers import AutoTokenizer, AutoModelForCausalLM

# Load model and tokenizer
tokenizer = AutoTokenizer.from_pretrained("VanshajR/gpt2-emotion-prefix")
model = AutoModelForCausalLM.from_pretrained("VanshajR/gpt2-emotion-prefix")

# Generate emotion-controlled response
emotion = "happy"
context = "How was your day today?"
prefix = f"Respond with {emotion} emotion: {context}"

inputs = tokenizer(prefix, return_tensors="pt")
outputs = model.generate(
    **inputs,
    max_length=100,
    num_return_sequences=1,
    temperature=0.8,
    top_p=0.9,
    do_sample=True,
    pad_token_id=tokenizer.eos_token_id
)

response = tokenizer.decode(outputs[0], skip_special_tokens=True)
print(response)
# Example: "Respond with happy emotion: How was your day today? It was amazing! I got promoted at work."

Supported Emotions

Best Practices

python

# For better results:
# 1. Use temperature 0.7-0.9 for natural responses
# 2. Use top_p=0.9 for diversity
# 3. Keep context concise (1-2 sentences)
# 4. Prefix format: "Respond with [emotion] emotion: [context]"

# Example with emotion verification
from transformers import pipeline

# Load emotion classifier for verification
classifier = pipeline("text-classification", model="VanshajR/roberta-emotion-7class")

# Generate with happy emotion
emotion = "happy"
context = "Tell me about your weekend"
prefix = f"Respond with {emotion} emotion: {context}"

inputs = tokenizer(prefix, return_tensors="pt")
outputs = model.generate(**inputs, max_length=80, temperature=0.8, top_p=0.9, do_sample=True)
response = tokenizer.decode(outputs[0], skip_special_tokens=True)

# Verify emotion
predicted = classifier(response)[0]
print(f"Target: {emotion}, Predicted: {predicted['label']}, Confidence: {predicted['score']:.2%}")

Training Details

Training Data

Dataset: DailyDialog
Training Samples: 76,052 multi-turn conversations
Test Samples: 6,740
Emotions: 7 classes (happy, sad, angry, fear, disgust, surprise, neutral)
Format: "Respond with [emotion] emotion: [dialogue_context]" → [response]

Training Procedure

Optimizer: AdamW (lr=5e-5, weight_decay=0.01)
Batch Size: 8 (with gradient accumulation=4, effective=32)
Epochs: 3
Max Length: 128 tokens
Training Regime: fp32
Gradient Clipping: 1.0

Compute Infrastructure

Hardware: NVIDIA RTX 3070 (8GB VRAM)
Training Time: ~8 hours (3 epochs)
Framework: PyTorch 2.1.0, Transformers 4.35.0

Method: Prefix Conditioning

This model uses natural language prefix conditioning - a lightweight method that doesn't modify the model architecture:

Baseline: "How are you?" → "I'm fine."
Prefix-Conditioned: "Respond with happy emotion: How are you?" → "I'm doing great! Everything's wonderful!"

Why Prefix > Token:

More interpretable (natural language)
Better performance (38.2% vs 30.8%)
No special tokens needed
Easier to use

Limitations and Considerations

Accuracy: 38.2% emotion match (not perfect control)
Evaluation: Automatic classifier (57.8% accuracy) - not human evaluation
Language: English conversational text only
Domain: Trained on daily conversations (may not work well for formal/technical text)
Size: Small model (124M) - less powerful than larger LLMs

What affects accuracy:

Classifier ceiling (57.8% max possible)
Task difficulty (humans disagree 25-30%)
Model size (124M vs billions in modern LLMs)
Training data size (76K vs millions)

Comparison with Literature

Paper	Model Size	Emotion Acc.	Dataset
Colombo et al. (2019)	117M	35-42%	DailyDialog
Zhou et al. (2018)	256M	~45%	DailyDialog
Rashkin et al. (2019)	345M-1.5B	40-48%	EmpatheticDialogues
This Work	124M	38.2%	DailyDialog

Our result falls within the published range for similar lightweight methods.

Intended Use

✅ Recommended:

Research on controllable text generation
Emotion-aware chatbot prototypes
Educational demonstrations of conditioning methods
Dialogue system experiments

❌ Not Recommended:

Production chatbots (use larger models)
Mental health applications
High-stakes communication
Non-English languages

Citation

bibtex

@misc{vanshajr2024gpt2emotion,
  author = {Vanshaj R},
  title = {GPT-2 Emotion-Conditioned Text Generation via Prefix Conditioning},
  year = {2024},
  publisher = {HuggingFace},
  url = {https://huggingface.co/VanshajR/gpt2-emotion-prefix}
}

Related Work

Part of the Emotion-Controlled Response Generation project:

Acknowledgments

DailyDialog Dataset: Li et al. (2017)
Base Model: OpenAI GPT-2
Evaluation: GoEmotions-trained RoBERTa classifier