README.md

metadata

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

NOTE: For experimental purposes

Chikuma is a 10.7B parameter model and is a merge of the following models using LazyMergekit:

• sethuiyer/SynthIQ-7b
• openchat/openchat-3.5-0106

The name "Chikuma" is inspired by the Chikuma River, the longest in Japan, known for its continuous flow and meandering path. This metaphorically represents the model's depth, fluidity, and adaptability in processing and understanding language.

It also perfectly fits the approach taken here - Depth Upscaling, inspired by SOLAR 10.7B.

Nous LLM Evaluation (Version 1 - with ChatML Prompt Template)

Model	AGIEval	GPT4All	TruthfulQA	Bigbench	Average
Chikuma_10.7B	42.41	73.41	56.69	43.5	54

More details can be found here

Recommended Prompt Template

<|im_start|>GPT4 Correct system
You are Chikuma, a constantly learning AI assistant who strives to be
insightful, engaging, and helpful. You possess vast knowledge and creativity,
but also a humble curiosity about the world and the people you interact
with. If you don't know the answer to a question, please don't share false information. 
Always use <|end_of_turn|> when you want to end the answer.<|im_end|>
<|im_start|>GPT4 Correct User:
{{Input}}
<|im_end|>GPT4 Correct Assistant:

ChatML format also works well.

Tested to work well in :

1. text-generation-webui, eos_token_id=32000, LLaMa-Precise sampling settings.
2. transformers text generation pipeline, temperature=4.0, top_k=50, top_p=0.01, eos_token_id=32000.

🧩 Configuration

slices:
  - sources:
    - model: sethuiyer/SynthIQ-7b
      layer_range: [0, 24]
  - sources:
    - model: openchat/openchat-3.5-0106
      layer_range: [8, 32]
merge_method: passthrough
dtype: bfloat16

💻 Usage

sys_message = ''' 
You are Chikuma, a constantly learning AI assistant who strives to be
insightful, engaging, and helpful. You possess vast knowledge and creativity,
but also a humble curiosity about the world and the people you interact
with. If you don't know the answer to a question, please don't share false information. 
Always use <|end_of_turn|> when you want to end the answer.
'''

question = '''
Tell me what is a large language model in under 250 words.
'''

messages = [{"role":"system", "content": sys_message}, {"role": "user", "content": question}]
prompt = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
outputs = pipeline(prompt, max_new_tokens=256, do_sample=True, temperature=4.0, top_k=50, top_p=0.01, eos_token_id=32000)
print(outputs[0]["generated_text"])