metadata

license: apache-2.0
language:
  - en
pipeline_tag: text-generation
library_name: transformers
tags:
  - llm
  - code

CrystalChat

We present CrystalChat, an instruction following model finetuned from LLM360/CrystalCoder. Following the release of LLM360/AmberChatand LLM360/AmberSafe in December 2023, CrystalChat is the next and most performant chat model released under LLM360. CrystalChat is trained on a carefully selected mix publicly available language and code datasets.

As always, the training data, training code, and metrics are publicly available.

About LLM360

LLM360 is an initiative for comprehensive and fully open-sourced LLMs, where all training details, model checkpoints, intermediate results, and additional analyses are made available to the community. Our goal is to advance the field by inviting the community to deepen the understanding of LLMs together. As the first step of the project LLM360, we release all intermediate model checkpoints, our fully-prepared pre-training dataset, all source code and configurations, and training details. We are committed to continually pushing the boundaries of LLMs through this open-source effort.

Get access now at LLM360 site

Instruction Tuning Training

CrystalChat is using the last CrystalCoder checkpoint of phase2 (CrystalCoder_phase2_checkpoint_214387) as the initialization checkpoint. We then finetune the model using the dataset mentioned below.

We also performed the same finetuning on the last CrystalCoder checkpoint of phase3 (CrystalCoder_phase3_checkpoint_027728). The phase2 and phase3 finetuning results are very similar, but phase2 finetuning exhibits slightly better performance on the English language benchmarks. We choose the phase2 finetuning result as the final model for CrystalChat.

Instruction Tuning Data

The instruction tuning data is a mix of publicly available language and code datasets, plus a orginally created dataset called WebAlpaca. The WebAlpaca dataset is created by us and is used as part of our instruction tuning training data. We will release the WebAlpaca dataset in a separate repository.

The summary of the instruction tuning data is as follows:

Instruction Format

We've added some new special tokens to the CrystalCoder tokenizer to support the instruction tuning.

List special tokens used in the instruction tuning:

bos: <s> 
eos: </s>
system_start: <|sys_start|>
system_end: <|sys_end|>
user_start: <|im_start|>
user_end: <|im_end|>

The instruction format is as follows:

<s> <|sys_start|> system prompt <|sys_end|> <|im_start|> first user utterance <|im_end|> first model response <|im_start|> next user utterance <|im_end|> next model response </s>

Reproducing the Results

We will realize the training code and the training data soon. Our training code is based on Megatron-LM, with some modifications to support our training data format and Maximal Update Parametrization (μP).

CrystalChat Performance

Model	Trained Tokens	Avg. of Avg.	Language Avg.	Coding Avg.	ARC	HellaSwag	MMLU (5-shot)	GSM8K	Winogrande(5-shot)	TruthfulQA	HumanEval (pass@1)	MBPP (pass@1)
CrystalChat 7B	1.275T	44.96	53.29	36.62	51.71	76.12	53.22	28.05	70.64	47.29	34.12	39.11
Mistral-7B-Instruct-v0.1	-	44.34	54.86	30.62	58.05	75.71	55.56	32.00	74.27	55.90	29.27	31.96
CodeLlama-7b-Instruct	2.5T	40.91	45.29	36.52	43.35	66.14	42.75	15.92	64.33	39.23	34.12	38.91
Llama-2-7b-Chat	2T	34.11	52.86	15.35	53.07	78.39	48.42	18.88	73.09	45.30	13.26	17.43
AmberChat 7B	1.25T	-	44.76	-	42.83	74.03	38.88	5.31	66.77	40.72	-	-

Combined Language and Coding Ability

Performance on Standard Benchmarks

Perforamnce on Language Benchmarks

Model Description

Model type: Language model with the same architecture as LLaMA-7B
Language(s) (NLP): English
License: Apache 2.0
Resources for more information:

Loading CrystalChat

import torch
from transformers import AutoModelForCausalLM, AutoTokenizer

device = "cuda:0" if torch.cuda.is_available() else "cpu"
tokenizer = AutoTokenizer.from_pretrained("LLM360/CrystalChat", trust_remote_code=True)
model = AutoModelForCausalLM.from_pretrained("LLM360/CrystalChat", trust_remote_code=True).to(device)

prompt = '<s> <|sys_start|> You are an AI assistant. You will be given a task. You must generate a detailed and long answer. <|sys_end|> <|im_start|> Write a python function that takes a list of integers and returns the squared sum of the list. <|im_end|>'


input_ids = tokenizer(prompt, return_tensors="pt").input_ids.to(device)
gen_tokens = model.generate(input_ids, do_sample=True, max_length=400)

print("-"*20 + "Output for model"  + 20 * '-')
print(tokenizer.batch_decode(gen_tokens)[0])

Response:

Here's a Python function named `squared_sum_list` that takes a list of integers as input and returns the squared sum of the list:

```python
def squared_sum_list(lst):
    return sum([num ** 2 for num in lst])
```

The function `squared_sum_list` uses a list comprehension to iterate over each number in the input list `lst` and calculate its square. Then, it uses the `sum` function to accumulate all the squared numbers in a single value - the squared sum of the list.

For example:

```python
numbers = [1, 2, 3, 4, 5]
print(squared_sum_list(numbers))  # Outputs: 55
```

In the above code, the list `[1, 2, 3, 4, 5]` is passed as an argument to the `squared_sum_list` function. The function calculates the sum of the squares of the elements in the list, which is `1 + 4 + 9 + 16 + 25 = 55`. The function then returns this result, which is printed to the console.</s>

Evaluation

Coming Soon!

Bias, Risks, and Limitations

CrystalChat has not been aligned to human preferences for safety within the RLHF phase or deployed with in-the-loop filtering of responses like ChatGPT, so the model can produce problematic outputs (especially when prompted to do so). The training data is known and made available here. It primarily consists of SlimPajama, StarCoder, and WebCrawl dataset.

Citation

BibTeX:

@misc{liu2023llm360,
      title={LLM360: Towards Fully Transparent Open-Source LLMs}, 
      author={Zhengzhong Liu and Aurick Qiao and Willie Neiswanger and Hongyi Wang and Bowen Tan and Tianhua Tao and Junbo Li and Yuqi Wang and Suqi Sun and Omkar Pangarkar and Richard Fan and Yi Gu and Victor Miller and Yonghao Zhuang and Guowei He and Haonan Li and Fajri Koto and Liping Tang and Nikhil Ranjan and Zhiqiang Shen and Xuguang Ren and Roberto Iriondo and Cun Mu and Zhiting Hu and Mark Schulze and Preslav Nakov and Tim Baldwin and Eric P. Xing},
      year={2023},
      eprint={2312.06550},
      archivePrefix={arXiv},
      primaryClass={cs.CL}
}