pints-ai/1.5-Pints-2K-v0.1

1.5-Pints -- A model pretrained in 9 days by using high quality data

Join us at Discord: https://discord.com/invite/RSHk22Z29j

Install dependencies

pip install transformers
# Omit `flash-attn` if not supported by your hardware
pip install flash-attn --no-build-isolation

Usage

from transformers import AutoModelForCausalLM, AutoTokenizer
device = "cuda" # the device to load the model onto

# INITIALIZE the model and the tokenizer
model = AutoModelForCausalLM.from_pretrained(
    "pints-ai/1.5-Pints-2k-v0.1",
    device_map=device,
    attn_implementation="flash_attention_2" # can be omitted if not supported
)
tokenizer = AutoTokenizer.from_pretrained("pints-ai/1.5-Pints-2k-v0.1")

# PREPARE and tokenize the prompt
prompt = "Predict what life will be like 100 years from now."
messages = [
    {"role": "system", "content": "You are an AI assistant that follows instruction extremely well. Help as much as you can."},
    {"role": "user", "content": prompt}
]

text = tokenizer.apply_chat_template(
    messages,
    tokenize=False,
    add_generation_prompt=True
)

input = tokenizer([text], return_tensors="pt").to(device)

# GENERATE the response
generated_ids = model.generate(
    input.input_ids,
    max_new_tokens=512
)

# DECODE the response
input_length = len(input.input_ids[0])
# Remove the input and decode only the output
response = tokenizer.decode(generated_ids[0][input_length:])

print(response)

Compute Infrastructure
This model can be served with a GPU containing at least 8GB of VRAM.

Description

1.5 Pints is a Large Language Model that significantly advances the efficiency of LLM training by emphasizing data quality over quantity. Our pre-training corpus is a meticulously curated dataset of 57 billion tokens, thus making pre-training more accessible and environmentally-friendly.

Results

MTBench
MTBench is a popular evaluation harness that uses strong LLMs like GPT-4 to act as judges and assess the quality of the models' responses.

Model	Score	Parameter Size	Pretrain Tokens
meta-llama/Llama-2-7b-chat-hf	6.27	7B	2T
microsoft/phi-2	5.83	2.7B	1.4T
google/gemma-2b-it	5.44	2B	3T
stabilityai/stablelm-2-1_6b-chat	4.7	1.6B	2T
1.5-Pints-2K	3.73	1.57B	0.115T
TinyLlama/TinyLlama-1.1B-Chat-v1.0	3.72	1.1B	3T
1.5-Pints-16K	3.40	1.57B	0.115T
apple/OpenELM-1_1B-Instruct	3.34	1B	1.8T
microsoft/phi-1_5	3.33	1.3B	0.15T
databricks/dolly-v2-3b	2.33	3B	0.3T
EleutherAI/pythia-2.8b	1.81	2.8B	0.3T
tiiuae/falcon-rw-1b	1.18	1B	0.35T

The 16K context window version of 1.5-Pints can be found here.

Technical Specifications

Architecture
Llama 2 Autoregressive Model with 2K Context Window and Mistral tokenizer. The model uses Float32 precision.

Parameters	Vocab Size	Embedding Size	Context Length	Layers	Heads	Query Groups	Intermediate Hidden Size
1,565,886,464	32,064	2,048	2,048	24	32	4	8,192

Context Lengths
1.5-Pints comes in 2 context lengths - 16k (16,384) and 2k (2,048).

Prompt template
This model has been finetuned and preference-optimized using the ChatML template.

<|im_start|>system 
{SYSTEM_PROMPT}<|im_end|> 
<|im_start|>user 
{PROMPT}<|im_end|> 
<|im_start|>assistant 

Uses

Direct Use
This model is meant to be an efficient and fine-tunable helpful assistant. It is designed to excel in user assistance and reasoning, and rely less on internal knowledge and factuals. Thus, for knowledge retrieval purposes, it should be used with Retrieval Augmented Generation.

Downstream Use
Given the size of this model, it is possible to launch multiple instances of it for use in agentic context without breaking the compute bank.

Recommendations

• It is recommended to finetune this model for domain adaption, and use it for a specialized tasks.
• To reap full performance, use a repetition penalty of 1.3 rather than 1.
  
  

Training Data

Pre-Train Data
Dataset: pints-ai/Expository-Prose-V1

Fine-Tune Data
Corpora:

• HuggingFaceH4/ultrachat
• Open-Orca/SlimOrca-Dedup
• meta-math/MetaMathQA
• HuggingFaceH4/deita-10k-v0-sft
• WizardLM/WizardLM_evol_instruct_V2_196k
• togethercomputer/llama-instruct
• LDJnr/Capybara

DPO Data
Dataset: HuggingFaceH4/ultrafeedback_binarized

Training Procedure

Both Pre-Train and Finetuning used our fork of the LitGPT Framework. For DPO, we used the methods set out in The Alignment Handbook. More details can be found in our paper.

Training Hyperparameters

Pre-Train

Hyperparameter	Value
Optimizer	AdamW(Beta1=0.9, Beta2=0.95)
Learning Rate Scheduler	Cosine
Max Learning Rate	4.0x10-4
Min Learning Rate	4.0x10-5
Warmup Steps	2,000
Batch Size	2,097,152
Weight Decay	0.1
Gradient Clipping Threshold	1.0

SFT

Hyperparameter	Value
Optimizer	AdamW(Beta1=0.9, Beta2=0.95)
Warmup steps	1,126 (10%)
Peak learning rate	2e-5
Learning rate scheduler	Cosine
Weight Decay	0.1

DPO
DPO parameters used are the exact same as those specified in The Alignment Handbook.

Citation

Attribution

• Developed by: calvintwr, lemousehunter
• Funded by PintsAI
• Released by: PintsAI
• Model type: Large Language Model
• Language(s) (NLP): English
• License: MIT License
  
  

BibTeX:

@misc{tan202415pintstechnicalreportpretraining,
      title={1.5-Pints Technical Report: Pretraining in Days, Not Months -- Your Language Model Thrives on Quality Data}, 
      author={Calvin Tan and Jerome Wang},
      year={2024},
      eprint={2408.03506},
      archivePrefix={arXiv},
      primaryClass={cs.CL},
      url={https://arxiv.org/abs/2408.03506}, 
}

APA
Tan, C., & Wang, J. (2024). 1.5-Pints Technical Report: Pretraining in days, not months -- Your language model thrives on quality data. arXiv. https://arxiv.org/abs/2408.03506

Legal Warning

Though best efforts has been made to ensure, as much as possible, that all texts in the training corpora are royalty free, this does not constitute a legal guarantee that such is the case. By using any of the models, corpora or part thereof, the user agrees to bear full responsibility to do the necessary due diligence to ensure that he / she is in compliance with their local copyright laws.

Additionally, the user agrees to bear any damages arising as a direct cause (or otherwise) of using any artifacts released by the pints research team, as well as full responsibility for the consequences of his / her usage (or implementation) of any such released artifacts. The user also indemnifies Pints Research Team (and any of its members or agents) of any damage, related or unrelated, to the release or subsequent usage of any findings, artifacts or code by the team.

For the avoidance of doubt, any artifacts released by the Pints Research team are done so in accordance with the "fair use" clause of Copyright Law, in hopes that this will aid the research community in bringing LLMs to the next frontier.