Edit model card

It remain factual and reliable even in dramatic situations.


Model Card for kevin009/llamaRAGdrama

Model Details

  • Model Name: kevin009/llamaRAGdrama
  • Model Type: Fine-tuned for Q&A, RAG.
  • Fine-tuning Objective: Synthesis text content in Q&A, RAG scenarios.

Intended Use

  • Applications: RAG, Q&A

Training Data

  • Sources: Includes a diverse dataset of dramatic texts, enriched with factual databases and reliable sources to train the model on generating content that remains true to real-world facts.
  • Preprocessing: In addition to removing non-content text, data was annotated to distinguish between purely creative elements and those that require factual accuracy, ensuring a balanced training approach.

How to Use

from transformers import AutoTokenizer, AutoModelForCausalLM

tokenizer = AutoTokenizer.from_pretrained("kevin009/llamaRAGdrama")
model = AutoModelForCausalLM.from_pretrained("kevin009/llamaRAGdrama")

input_text = "Enter your prompt here"
input_tokens = tokenizer.encode(input_text, return_tensors='pt')
output_tokens = model.generate(input_tokens, max_length=100, num_return_sequences=1, temperature=0.9)
generated_text = tokenizer.decode(output_tokens[0], skip_special_tokens=True)

print(generated_text)

Replace "Enter your prompt here" with your starting text. Adjust temperature for creativity level.

Limitations and Biases

  • Content Limitation: While designed to be truthful, It may not be considered safe.
  • Biases: It may remain biases and inaccurate.

Licensing and Attribution

  • License: Apache-2.0

Open LLM Leaderboard Evaluation Results

Detailed results can be found here

Metric Value
Avg. 74.65
AI2 Reasoning Challenge (25-Shot) 72.01
HellaSwag (10-Shot) 88.83
MMLU (5-Shot) 64.50
TruthfulQA (0-shot) 70.24
Winogrande (5-shot) 86.66
GSM8k (5-shot) 65.66
Downloads last month
5,048
Safetensors
Model size
7.24B params
Tensor type
FP16
Β·
Inference Examples
This model does not have enough activity to be deployed to Inference API (serverless) yet. Increase its social visibility and check back later, or deploy to Inference Endpoints (dedicated) instead.

Model tree for kevin009/llamaRAGdrama

Quantizations
2 models

Spaces using kevin009/llamaRAGdrama 6

Evaluation results