Mistral-Nemo-NT-Ko-12B-sft

Mistral-Nemo-NT-Ko-12B-sft

Description

Mistral-Nemo-NT-Ko-12B-sft is an instruction-tuned version of mistralai/Mistral-Nemo-Base-2407, fine-tuned across four languages: English, Korean, Chinese, and Japanese.

The primary goals of this model are language alignment, cross-lingual knowledge transfer and ChatML formatting. This is an intermediate version since preference optimization has not yet been applied.

Features

• The base model supports a context length of 128K, while I fine-tuned this model with an 8K context size.

• The model follows to the input language unless the user explicitly specifies an output language (If the language is set by a system role, it may be ignored).

• Answer length tends to vary by language: English responses are generally longer than average, while Korean responses tend to be shorter. The behavior for Japanese and Chinese is still under observation.

• Recommended temperature settings: 0.3 to 0.7.

Evaluation

LogicKor

MT-Bench

Language-Confusion(Korean Only)

example:

<|im_start|>system
You are a helpful AI assistant.<|im_end|>
<|im_start|>user
{prompt}<|im_end|>
<|im_start|>assistant

I trained Mistral-Nemo-NT-Ko-12B with various system prompt from dozens of dataset. You can chat with/without your system prompt.

Dataset

werty1248/multilingual-instruct-balanced

Training Details

• GPU: 8xA40
• epoch: 3
• total batch size: 8
• learning rate: 7e-6
• weight decay: 0.01

<img src="https://raw.githubusercontent.com/axolotl-ai-cloud/axolotl/main/image/axolotl-badge-web.png" alt="Built with Axolotl" width="200" height="32"/>

<details><summary>See axolotl config</summary>

axolotl version: 0.4.1

base_model: mistralai/Mistral-Nemo-Base-2407
model_type: MistralForCausalLM
tokenizer_config: nothingiisreal/MN-12B-Celeste-V1.9 ##axolotl-ai-co/Mistral-Nemo-Base-2407-chatml makes error, why?
tokenizer_type: AutoTokenizer

load_in_8bit: false
load_in_4bit: false
strict: false

chat_template: chatml
datasets:
  - path: werty1248/multilingual-instruct-balanced
    type: sharegpt
    chat_template: chatml

dataset_prepared_path: ./data_preparation
output_dir: /workspace/data

hf_use_auth_token: true

sequence_len: 8192
sample_packing: true
pad_to_sequence_len: true

wandb_project:
#wandb_entity:
#wandb_watch:
wandb_name:
#wandb_log_model:

gradient_accumulation_steps: 1 ## total_batch = 8
micro_batch_size: 1
num_epochs: 3
optimizer: paged_adamw_32bit
lr_scheduler: cosine
learning_rate: 0.000007

train_on_inputs: false
group_by_length: false
bf16: auto
fp16: 
tf32: false

gradient_checkpointing: true
early_stopping_patience:
resume_from_checkpoint:
local_rank:
logging_steps: 1
xformers_attention:
flash_attention: true

warmup_steps: 1000
evals_per_epoch: 1
eval_table_size:
save_steps: 1000
debug:
deepspeed: deepspeed_configs/zero3_bf16.json
weight_decay: 0.01
special_tokens:
  pad_token: <pad>

</details><br>

• Training loss