This model is a fine-tuned version of NousResearch/Llama-2-7b-chat-hf on the GoEmotions dataset with activity context integration. It analyzes emotions in text while considering the user's recent activity patterns to provide more contextual insights.

Training Details

Training Data

Dataset: AA65327/GoEmotions_Alpaca_Final
Training samples: N/A
Validation samples: N/A

Training Configuration

Base model: NousResearch/Llama-2-7b-chat-hf
Training epochs: 1
Batch size: 1
Learning rate: 0.0002
LoRA rank: 8
LoRA alpha: 32

Performance Metrics

Evaluation Results

Perplexity: 26.08
ROUGE-1: 0.190
ROUGE-2: 0.170
ROUGE-L: 0.190
BLEU Score: 8.039
Inference Speed: 1.3 tokens/sec
Hallucination Rate: 2.400

Usage

python

from transformers import AutoTokenizer, AutoModelForCausalLM
from peft import PeftModel

# Load model and tokenizer
base_model = AutoModelForCausalLM.from_pretrained(
    "NousResearch/Llama-2-7b-chat-hf",
    load_in_4bit=True,
    device_map="auto"
)
model = PeftModel.from_pretrained(base_model, "AA65327/llama2-emotion-activity-20251005")
tokenizer = AutoTokenizer.from_pretrained("AA65327/llama2-emotion-activity-20251005")

# Format your prompt
def format_prompt(instruction, input_text, activity_log):
    return f"""Below is an instruction that describes a task, paired with an input that provides further context.
Write a response that appropriately completes the request.

### Instruction:
{instruction}

### Input:
Current message: {input_text}
Activity log (past 3 days, hours per activity): {activity_log}

### Response:
"""

# Example usage
instruction = "Evaluate the emotion in this text and suggest why the person might feel this way."
input_text = "I'm feeling really excited about this new project!"
activity_log = "working_out: [2, 1, 3]; reading: [1, 2, 0]; socializing: [3, 4, 2]"

prompt = format_prompt(instruction, input_text, activity_log)
inputs = tokenizer(prompt, return_tensors="pt")
outputs = model.generate(**inputs, max_new_tokens=150, temperature=0.7)
response = tokenizer.decode(outputs[0], skip_special_tokens=True)
print(response)

Training Procedure

The model was trained using LoRA (Low-Rank Adaptation) technique with the following approach:

Load base LLaMA-2-7B-Chat model with 4-bit quantization
Apply LoRA adapters to query and value projection layers
Fine-tune on emotion analysis tasks with activity context
Implement gradient checkpointing and mixed precision training
Use early stopping based on validation loss

Limitations and Bias

The model may reflect biases present in the training data
Performance may vary on domains not represented in the training set
Activity context interpretation is based on patterns learned from training data
Generated content should be reviewed for factual accuracy

Citation

bibtex

@misc{llama2-emotion-activity-2025,
  author = {AA65327},
  title = {LLaMA-2-7B Emotion Analysis with Activity Context},
  year = {2025},
  publisher = {Hugging Face},
  url = {https://huggingface.co/AA65327/llama2-emotion-activity-20251005}
}

Acknowledgments

Meta AI for the base LLaMA-2 model
Google Research for the GoEmotions dataset
Hugging Face for the transformers library and model hosting