Kendamarron/Qwen2.5-1.75B-A1.1B-Instruct-ja

Qwen2.5-1.75B-A1.1B-Instruct-ja

Qwen2.5-0.5B系のモデルを組み合わせて作ったMoEです。

Details

https://zenn.dev/kendama/articles/68ae234e9371ac

See axolotl config

axolotl version: 0.6.0

# 学習のベースモデルに関する設定
base_model: Kendamarron/Qwen2.5-4x0.5B-cpt
model_type: AutoModelForCausalLM
tokenizer_type: AutoTokenizer

# 学習後のモデルのHFへのアップロードに関する設定
hub_model_id: Kendamarron/Qwen2.5-4x0.5B-sft-v1
hub_strategy: "end"
push_dataset_to_hub:
hf_use_auth_token: true

# Liger Kernelの設定（学習の軽量・高速化）
plugins:
  - axolotl.integrations.liger.LigerPlugin
liger_cross_entropy: false
liger_rope: true
liger_rms_norm: true
liger_swiglu: true
liger_fused_linear_cross_entropy: true

# 量子化に関する設定
load_in_8bit: false
load_in_4bit: false

# SFTに利用するchat templateの設定
chat_template: qwen_25

# 学習データセットの前処理に関する設定
datasets:
  - path: Kendamarron/jimba-instruction-all
    split: train
    type: chat_template
    field_messages: conversations
    message_field_role: role
    message_field_content: content
  - path: Kendamarron/OpenMathInstruct-2-ja-CoT-only_thought
    split: train
    type: chat_template
    field_messages: messages
    message_field_role: role
    message_field_content: content
  - path: Aratako/Synthetic-JP-EN-Coding-Dataset-801k
    split: train[0:10000]
    type: chat_template
    field_messages: messages
    message_field_role: role
    message_field_content: content
  - path: llm-jp/magpie-sft-v1.0
    split: train[0:30000]
    type: chat_template
    field_messages: conversations
    message_field_role: role
    message_field_content: content


# データセット、モデルの出力先に関する設定
shuffle_merged_datasets: true
dataset_prepared_path: /workspace/data/sft-data
output_dir: /workspace/data/models/Qwen2.5-4x0.5B-SFT

# valid datasetのサイズ
val_set_size: 0.005

# wandbに関する設定
wandb_project: Qwen2.5-4x0.5B
wandb_entity: kendamarron
wandb_watch:
wandb_name: sft-v1
wandb_log_model:

# 学習に関する様々な設定
sequence_len: 8192
sample_packing: true
eval_sample_packing: false
pad_to_sequence_len: true

gradient_accumulation_steps: 8
micro_batch_size: 1
num_epochs: 2
optimizer: adamw_bnb_8bit
lr_scheduler: cosine
cosine_min_lr_ratio: 0.1
learning_rate: 2e-5

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

gradient_checkpointing: false
early_stopping_patience:
auto_resume_from_checkpoints: true
local_rank:
logging_steps: 1
xformers_attention:
flash_attention: true

saves_per_epoch: 1

warmup_steps: 60
eval_steps: 100
eval_batch_size: 1
eval_table_size:
eval_max_new_tokens:
debug:
deepspeed: deepspeed_configs/zero3_bf16.json
weight_decay: 0.1
fsdp:
fsdp_config:
special_tokens:
  eos_token: "<|im_end|>"
  pad_token: "<|end_of_text|>"
tokens:
  - "<|im_start|>"
  - "<|im_end|>"

Qwen2.5-4x0.5B-sft-v1

This model is a fine-tuned version of Kendamarron/Qwen2.5-4x0.5B-cpt on the Kendamarron/jimba-instruction-all, the Kendamarron/OpenMathInstruct-2-ja-CoT-only_thought, the Aratako/Synthetic-JP-EN-Coding-Dataset-801k and the llm-jp/magpie-sft-v1.0 datasets. It achieves the following results on the evaluation set:

• Loss: 1.0085

Model description

More information needed

Intended uses & limitations

More information needed

Training and evaluation data

More information needed

Training procedure

Training hyperparameters

The following hyperparameters were used during training:

• learning_rate: 2e-05
• train_batch_size: 1
• eval_batch_size: 1
• seed: 42
• distributed_type: multi-GPU
• num_devices: 4
• gradient_accumulation_steps: 8
• total_train_batch_size: 32
• total_eval_batch_size: 4
• optimizer: Use OptimizerNames.ADAMW_BNB with betas=(0.9,0.999) and epsilon=1e-08 and optimizer_args=No additional optimizer arguments
• lr_scheduler_type: cosine
• lr_scheduler_warmup_steps: 60
• num_epochs: 2

Training results

Training Loss	Epoch	Step	Validation Loss
1.3068	0.0033	1	1.3071
1.1087	0.3309	100	1.0806
1.1393	0.6617	200	1.0488
1.0569	0.9926	300	1.0286
0.9902	1.3209	400	1.0215
0.9933	1.6518	500	1.0133
0.9706	1.9826	600	1.0085

Framework versions

• Transformers 4.47.1
• Pytorch 2.5.1+cu124
• Datasets 3.1.0
• Tokenizers 0.21.0