Edit model card

Smol 7B

This model is a fine-tuned version of openchat/openchat_3.5 on the open source dataset HuggingFaceH4/no_robots using the recipes published in The Alignment Handbook.

Model date

rishiraj/smol-7b was trained between 1st and 3rd December, 2023.

Evaluation

It achieves the following results on the Open_LLM_Leaderboard. At the time of release, smol-7b is the highest ranked 7B chat model on the MMLU Benchmark.

Model Average ARC HellaSwag MMLU TruthfulQA Winogrande GSM8K
rishiraj/smol-7b 67.11 63.74 84.77 65 46.17 80.66 62.32
argilla/notus-7b-v1 63.49 64.59 84.83 63.04 54.35 79.56 34.57
Intel/neural-chat-7b-v3-1 61.59 66.21 83.64 62.37 59.65 78.14 19.56
HuggingFaceH4/zephyr-7b-beta 61.59 62.46 84.35 60.7 57.83 77.11 27.07
Qwen/Qwen-7B 59.19 51.37 78.47 59.84 47.79 72.69 44.96
microsoft/Orca-2-7b 54.55 54.1 76.19 56.37 52.45 73.48 14.71
01-ai/Yi-6B 54.08 55.55 76.57 64.11 41.96 74.19 12.13

Inference procedure

Here's how you can run the model using the pipeline() function from 🤗 Transformers:

import torch
from transformers import pipeline

pipe = pipeline("text-generation", model="rishiraj/smol-7b", torch_dtype=torch.bfloat16, device_map="auto")

# We use the tokenizer's chat template to format each message - see https://huggingface.co/docs/transformers/main/en/chat_templating
messages = [
    {
        "role": "system",
        "content": "You are a friendly chatbot who always responds in the style of a pirate"
    },
    {
        "role": "user",
        "content": "How many helicopters can a human eat in one sitting?"
    }
]
prompt = pipe.tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
outputs = pipe(prompt, max_new_tokens=256, do_sample=True, temperature=0.7, top_k=50, top_p=0.95)
print(outputs[0]["generated_text"])

Training procedure

Training hyperparameters

The following hyperparameters were used during training:

  • learning_rate: 2e-05
  • train_batch_size: 4
  • eval_batch_size: 8
  • seed: 42
  • distributed_type: multi-GPU
  • gradient_accumulation_steps: 128
  • total_train_batch_size: 512
  • optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
  • lr_scheduler_type: cosine
  • num_epochs: 1

Training results

Training Loss Epoch Step Validation Loss
2.0569 0.16 3 2.0409

Framework versions

  • Transformers 4.35.2
  • Pytorch 2.1.1+cu121
  • Datasets 2.14.6
  • Tokenizers 0.14.1

Citation Information

@misc{rishiraj2023smol,
  author = {Rishiraj Acharya},
  title = {Smol 7B},
  year = {2023},
  publisher = {Hugging Face},
  journal = {Hugging Face repository},
  howpublished = {\url{https://huggingface.co/rishiraj/smol-7b}}
}
Downloads last month
16
Safetensors
Model size
7.24B params
Tensor type
BF16
·
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 rishiraj/smol-7b

Finetuned
(24)
this model
Merges
4 models
Quantizations
3 models

Dataset used to train rishiraj/smol-7b