DeepSeek-R1-Distill-Qwen-32B-axolotl-sft-v1.0

<!-- This model card has been generated automatically according to the information the Trainer had access to. You should probably proofread and complete it, then remove this comment. -->

<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: 0.8.0.dev0

# 学習のベースモデルに関する設定
base_model: deepseek-ai/DeepSeek-R1-Distill-Qwen-32B
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: kazuyamaa/DeepSeek-R1-Distill-Qwen-32B-axolotl-sft-v1.0
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: true

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

# 学習データセットの前処理に関する設定
datasets:
  - path: kanhatakeyama/ramdom-to-fixed-multiturn-Calm3
    split: 20240806filtered[0:10000]
    type: chat_template
    field_messages: messages
    message_field_role: role
    message_field_content: content
  - path: Aratako/Magpie-Tanuki-Qwen2.5-72B-Answered
    split: train[0:10000]
    type: chat_template
    field_messages: messages
    message_field_role: role
    message_field_content: content
  - path: Aratako/magpie-qwen2.5-32b-reasoning-100k-formatted
    split: train[0:10000]
    type: chat_template
    field_messages: conversations
    message_field_role: role
    message_field_content: content
  - path: Aratako/magpie-reasoning-llama-nemotron-70b-100k-filtered
    split: train[0:10000]
    type: chat_template
    field_messages: conversations
    message_field_role: role
    message_field_content: content
  - path: Aratako/Open-Platypus-Japanese-masked-formatted
    split: train[0:10000]
    type: chat_template
    field_messages: conversations
    message_field_role: role
    message_field_content: content
  - path: kanhatakeyama/wizardlm8x22b-logical-math-coding-sft_additional-ja
    split: train[0:10000]
    type: chat_template
    field_messages: messages
    message_field_role: role
    message_field_content: content
  - path: Aratako/magpie-ultra-v0.1-formatted
    split: train[0:10000]
    type: chat_template
    field_messages: conversations
    message_field_role: role
    message_field_content: content
  - path: Aratako/orca-agentinstruct-1M-v1-selected
    split: train[0:10000]
    type: chat_template
    field_messages: messages
    message_field_role: role
    message_field_content: content
  - path: Aratako/Synthetic-JP-EN-Coding-Dataset-801k-50k
    split: train[0:10000]
    type: chat_template
    field_messages: messages
    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: /content/output/DeepSeek-R1-Distill-Qwen-32B-axolotl-sft-v1.0

# valid datasetのサイズ
val_set_size: 0.05

# LoRAに関する設定（フルファインチューニングしたい場合は全て空欄にする）
adapter: qlora
lora_model_dir:
lora_r: 16
lora_alpha: 32
lora_dropout: 0.05
lora_target_linear: true
lora_fan_in_fan_out:

# wandbに関する設定
#wandb_project: axolotl
#wandb_entity: kazukitakayamas051
#wandb_watch:
#wandb_name: sft-lora-1
#wandb_log_model:

# 学習に関する様々な設定
sequence_len: 4096
sample_packing: true
eval_sample_packing: false
pad_to_sequence_len: true

gradient_accumulation_steps: 16
micro_batch_size: 1
num_epochs: 1
optimizer: paged_adamw_8bit
lr_scheduler: cosine
cosine_min_lr_ratio: 0.1
learning_rate: 3e-4

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

save_strategy: steps
save_steps: 50
save_total_limit: 2

warmup_steps: 10
eval_steps: 50
eval_batch_size: 1
eval_table_size:
eval_max_new_tokens:
debug:
deepspeed: /workspace/axolotl/deepspeed_configs/zero3_bf16.json
weight_decay: 0.01
fsdp:
fsdp_config:
special_tokens:
  pad_token: <pad>

</details><br>

DeepSeek-R1-Distill-Qwen-32B-axolotl-sft-v1.0

This model is a fine-tuned version of deepseek-ai/DeepSeek-R1-Distill-Qwen-32B on the kanhatakeyama/ramdom-to-fixed-multiturn-Calm3, the Aratako/Magpie-Tanuki-Qwen2.5-72B-Answered, the Aratako/magpie-qwen2.5-32b-reasoning-100k-formatted, the Aratako/magpie-reasoning-llama-nemotron-70b-100k-filtered, the Aratako/Open-Platypus-Japanese-masked-formatted, the kanhatakeyama/wizardlm8x22b-logical-math-coding-sft_additional-ja, the Aratako/magpie-ultra-v0.1-formatted, the Aratako/orca-agentinstruct-1M-v1-selected and the Aratako/Synthetic-JP-EN-Coding-Dataset-801k-50k datasets. It achieves the following results on the evaluation set:

• Loss: 0.6154

以下、Axolotlの実行コード

!apt-get update
!apt-get install -y libopenmpi-dev

!git clone https://github.com/axolotl-ai-cloud/axolotl

cd axolotl
!pip install -e .
!pip install packaging ninja
!pip install flash-attn
!pip install deepspeed
!pip install mpi4py

# write権限のあるtokenを利用してHFにログイン（学習後のモデルアップロードに必要）
!huggingface-cli login --token WRITE ME
# wandbにログイン（wandbに学習ログを残したい場合）
!wandb login WRITE ME

import axolotl

!python -m axolotl.cli.preprocess /workspace/deepseek-32b-ver001-simpo.yml --debug

! accelerate launch -m axolotl.cli.train /workspace/deepseek-32b-ver001-simpo.yml --deepspeed deepspeed_configs/zero3_bf16.json

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: 0.0003
• train_batch_size: 1
• eval_batch_size: 1
• seed: 42
• distributed_type: multi-GPU
• gradient_accumulation_steps: 16
• total_train_batch_size: 16
• optimizer: Use paged_adamw_8bit 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: 10
• num_epochs: 1.0

Training results

Framework versions

• PEFT 0.14.0
• Transformers 4.49.0
• Pytorch 2.5.1+cu124
• Datasets 3.2.0
• Tokenizers 0.21.0