rishiraj/CatPPT-base
48271
1---2license: apache-2.03tags:4- merge5---6 7# 😼 CatPPT8 9Introducing "CatPPT" - the purrfect alternative to that other big cat in town, known for keeping all the secrets to itself! Our feline friend here is created through merging openchat and neuralchat models using Gradient SLERP method (resulting in [rishiraj/CatPPT-base](https://huggingface.co/rishiraj/CatPPT-base)) and then finetuned on no_robots dataset for chat.10 11This is the top-performing 7B model on the leaderboard, that's free from any whiff of evaluation data contamination.12 1314 15## Model date16 17rishiraj/CatPPT was trained between 15th and 17th December, 2023.18 19## Evaluation20 21It achieves the following results on the [Open_LLM_Leaderboard](https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard). At the time of release, CatPPT is the highest ranked 7B chat model on the leaderboard, that's **free from evaluation data contamination**.22 23|Model |Average|ARC |HellaSwag|MMLU |TruthfulQA|Winogrande|GSM8K|24|------------------------------------|-------|-----|---------|-----|----------|----------|-----|25|**rishiraj/CatPPT** |**72.32** |**68.09**|**86.69** |**65.16**|**61.55** |**81.61** |**70.81**|26|Intel/neural-chat-7b-v3-3 |69.83 |66.89|85.26 |63.07|63.01 |79.64 |61.11|27|openchat/openchat-3.5-1210 |68.89 |64.93|84.92 |64.62|52.15 |80.74 |65.96|28|meta-math/MetaMath-Mistral-7B |65.78 |60.67|82.58 |61.95|44.89 |75.77 |68.84|29|Deci/DeciLM-7B-instruct |63.19 |61.01|82.37 |60.24|49.75 |79.72 |46.02|30|mistralai/Mistral-7B-Instruct-v0.2 |65.71 |63.14|84.88 |60.78|68.26 |77.19 |40.03|31|mistralai/Mixtral-8x7B-Instruct-v0.1|72.62 |70.22|87.63 |71.16|64.58 |81.37 |60.73|32|meta-llama/Llama-2-70b-hf |67.87 |67.32|87.33 |69.83|44.92 |83.74 |54.06|33|tiiuae/falcon-180B |67.85 |69.45|88.86 |70.5 |45.47 |86.9 |45.94|34 35## Inference procedure36 37Here's how you can run the model using the pipeline() function from 🤗 Transformers:38 39```40import torch41from transformers import pipeline42 43pipe = pipeline("text-generation", model="rishiraj/CatPPT", torch_dtype=torch.bfloat16, device_map="auto")44 45# We use the tokenizer's chat template to format each message - see https://huggingface.co/docs/transformers/main/en/chat_templating46messages = [47 {48 "role": "system",49 "content": "You are a friendly chatbot who always responds in the style of a pirate"50 },51 {52 "role": "user",53 "content": "How many helicopters can a human eat in one sitting?"54 }55]56prompt = pipe.tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)57outputs = pipe(prompt, max_new_tokens=256, do_sample=True, temperature=0.7, top_k=50, top_p=0.95)58print(outputs[0]["generated_text"])59```60 61## Training procedure62 63### Training hyperparameters64 65The following hyperparameters were used during training:66- learning_rate: 2e-0567- train_batch_size: 468- eval_batch_size: 869- seed: 4270- distributed_type: multi-GPU71- gradient_accumulation_steps: 12872- total_train_batch_size: 51273- optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-0874- lr_scheduler_type: cosine75- num_epochs: 176 77### Training results78 79| Training Loss | Epoch | Step | Validation Loss |80|:-------------:|:-----:|:----:|:---------------:|81| 1.9947 | 0.16 | 3 | 2.0093 |82 83 84### Framework versions85 86- Transformers 4.36.187- Pytorch 2.1.2+cu12188- Datasets 2.14.689- Tokenizers 0.15.090- PEFT 0.6.191 92## Citation Information93 94```95@misc{rishiraj2023catppt,96 author = {Rishiraj Acharya},97 title = {CatPPT},98 year = {2023},99 publisher = {Hugging Face},100 journal = {Hugging Face repository},101 howpublished = {\url{https://huggingface.co/rishiraj/CatPPT}}102}103```