Ninja-v1-128k / README.md
umisetokikaze's picture
Update README.md
6ccd171 verified
metadata
license: apache-2.0
language:
  - en
  - ja
tags:
  - finetuned
library_name: transformers
pipeline_tag: text-generation

Our Models

Model Card for Ninja-v1-128k

The Mistral-7B--based Large Language Model (LLM) is an noveldataset fine-tuned version of the Mistral-7B-v0.1

Ninja-128k has the following changes compared to Mistral-7B-v0.1.

  • 128k context window (8k context in v0.1)
  • Achieving both high quality Japanese and English generation
  • Memory ability that does not forget even after long-context generation

This model was created with the help of GPUs from the first LocalAI hackathon.

We would like to take this opportunity to thank

List of Creation Methods

  • Chatvector for multiple models
  • Simple linear merging of result models
  • Domain and Sentence Enhancement with LORA
  • Context expansion

Instruction format

Ninja adopts the prompt format from Vicuna and supports multi-turn conversation. The prompt should be as following:

USER: Hi ASSISTANT: Hello.</s>
USER: Who are you?
ASSISTANT: I am ninja.</s>

Example prompts to improve (Japanese)

  • BAD:ใ€€ใ‚ใชใŸใฏโ—‹โ—‹ใจใ—ใฆๆŒฏใ‚‹่ˆžใ„ใพใ™

  • GOOD: ใ‚ใชใŸใฏโ—‹โ—‹ใงใ™

  • BAD: ใ‚ใชใŸใฏโ—‹โ—‹ใŒใงใใพใ™

  • GOOD: ใ‚ใชใŸใฏโ—‹โ—‹ใ‚’ใ—ใพใ™

Performing inference

from transformers import AutoModelForCausalLM, AutoTokenizer
import torch

model_id = "Local-Novel-LLM-project/Ninja-v1-128k"
new_tokens = 1024

model = AutoModelForCausalLM.from_pretrained(model_id, trust_remote_code=True, torch_dtype=torch.float16, attn_implementation="flash_attention_2", device_map="auto")
tokenizer = AutoTokenizer.from_pretrained(model_id)

system_prompt = "ใ‚ใชใŸใฏใƒ—ใƒญใฎๅฐ่ชฌๅฎถใงใ™ใ€‚\nๅฐ่ชฌใ‚’ๆ›ธใ„ใฆใใ ใ•ใ„\n-------- "

prompt = input("Enter a prompt: ")
system_prompt += prompt + "\n-------- "
model_inputs = tokenizer([system_prompt], return_tensors="pt").to("cuda")


generated_ids = model.generate(**model_inputs, max_new_tokens=new_tokens, do_sample=True)
print(tokenizer.batch_decode(generated_ids)[0])

Merge recipe

  • WizardLM2 - mistralai/Mistral-7B-v0.1
  • NousResearch/Yarn-Mistral-7b-128k - mistralai/Mistral-7B-v0.1
  • Elizezen/Antler-7B - stabilityai/japanese-stablelm-instruct-gamma-7b
  • NTQAI/chatntq-ja-7b-v1.0

The characteristics of each model are as follows.

  • WizardLM2: High quality multitasking model
  • Yarn-Mistral-7b-128k: Mistral model with 128k context window
  • Antler-7B: Model specialized for novel writing
  • NTQAI/chatntq-ja-7b-v1.0 High quality Japanese specialized model

Other points to keep in mind

  • The training data may be biased. Be careful with the generated sentences.
  • Set trust_remote_code to True for context expansion with YaRN.
  • Memory usage may be large for long inferences.
  • If possible, we recommend inferring with llamacpp rather than Transformers.