See axolotl config

axolotl version: 0.10.0.dev0

# === Start-up Commands ===
# curl -LsSf https://astral.sh/uv/install.sh | sh
# export PATH="$HOME/.local/bin:$PATH"
# uv venv
# source .venv/bin/activate
# git clone https://github.com/axolotl-ai-cloud/axolotl
# cd axolotl
# uv pip install torch==2.5.1 packaging ninja setuptools ftfy deepspeed huggingface_hub[cli,hf_transfer]
# uv pip install "cut-cross-entropy[transformers] @ git+https://github.com/strangedove/ml-cross-entropy.git@gemma3-multimodal"
# uv pip install apollo-torch
# uv pip install --no-build-isolation -e .[flash-attn]
# uv pip install git+https://github.com/huggingface/transformers.git
# uv pip install git+https://github.com/linkedin/Liger-Kernel.git
# export HF_HUB_ENABLE_HF_TRANSFER=1
# huggingface-cli login --token $hf_key && wandb login $wandb_key

# apt update && apt install -y libopenmpi-dev && curl -LsSf https://astral.sh/uv/install.sh | sh && export PATH="$HOME/.local/bin:$PATH" && git clone https://github.com/axolotl-ai-cloud/axolotl && uv venv && source .venv/bin/activate && cd axolotl && uv pip install torch==2.5.1 packaging ninja mpi4py setuptools ftfy deepspeed huggingface_hub[cli,hf_transfer] && uv pip install apollo-torch && uv pip install "cut-cross-entropy[transformers] @ git+https://github.com/strangedove/ml-cross-entropy.git@qwen3" && uv pip install git+https://github.com/linkedin/Liger-Kernel.git && uv pip install --no-build-isolation -e .[flash-attn] && uv pip install git+https://github.com/huggingface/transformers.git && export HF_HUB_ENABLE_HF_TRANSFER=1 && cd .. && huggingface-cli login --token $hf_key && wandb login $wandb_key

# === Model Configuration ===
base_model: Columbidae/Qwen3-16B-A3B-Base
load_in_8bit: false
load_in_4bit: false

# === HF Configuration === 
hub_model_id: Columbidae/Qwen3-16B-A3B-Tulu-Mini
hub_strategy: "every_save"

# === Training Setup ===
num_epochs: 1
micro_batch_size: 4
gradient_accumulation_steps: 1
sequence_len: 4096
sample_packing: true
pad_to_sequence_len: true

# === Evaluation ===
val_set_size: 1000
evals_per_epoch: 5
#eval_steps: 20
#max_steps: 60
#eval_table_size:
eval_max_new_tokens: 256
eval_sample_packing: true
#eval_strategy: "no"

# === LoRA Configuration ===
#adapter: lora
#lora_model_dir:
#lora_r: 32
#lora_alpha: 32
#lora_dropout: 0
#lora_target_linear: 
#lora_fan_in_fan_out:
#lora_target_modules:
#  - gate_proj
#  - down_proj
#  - up_proj
#  - q_proj
#  - v_proj
#  - k_proj
#  - o_proj

#lora_mlp_kernel: true
#lora_qkv_kernel: true
#lora_o_kernel: true

# === Hyperparameter Configuration ===
optimizer: apollo_adamw_layerwise
#optimizer: paged_adamw_8bit
# Apollo-mini configuration:
optim_args: "proj=random,rank=128,scale=128.0,scale_type=tensor,update_proj_gap=100"
# Regular Apollo configuration:
# optim_args: 
optim_target_modules: all_linear
learning_rate: 3e-5
lr_scheduler: cosine
#lr_scheduler: cosine_with_min_lr
#lr_scheduler_kwargs:
#  cosine_min_lr: 1e-6
weight_decay: 0.01
#warmup_steps: 0
warmup_ratio: 0.025


# === Data Configuration ===
#chat_template: jinja
#chat_template_jinja: "{{ bos_token }}{% for message in messages %}{% if (message['role'] == 'assistant') %}{% set role = 'model' %}{% else %}{% set role = message['role'] %}{% endif %}{{ '<start_of_turn>' + role + '\n' + message['content'] | trim + '<end_of_turn>\n' }}{% endfor %}{% if add_generation_prompt %}{{'<start_of_turn>model\n'}}{% endif %}"
#special_tokens:
#  eos_token: "<end_of_turn>"
chat_template: chatml
shuffle_merged_datasets: true
datasets:
  - path: ToastyPigeon/tulu-mini
    type: chat_template
    
dataset_prepared_path: last_run_prepared


# === Plugins ===
plugins:
  - axolotl.integrations.liger.LigerPlugin
  - axolotl.integrations.cut_cross_entropy.CutCrossEntropyPlugin

# === Hardware Optimization ===
gradient_checkpointing: offload
gradient_checkpointing_kwargs:
  use_reentrant: false
liger_rope: true
liger_rms_norm: true
liger_glu_activation: true
#liger_fused_linear_cross_entropy: true
#unsloth_cross_entropy_loss: true
cut_cross_entropy: true
# Only if using multiple GPUs:
#deepspeed: axolotl/deepspeed_configs/zero2.json

# === Wandb Tracking ===
wandb_project: Qwen3MoE-Apollo
# wandb_entity: [WANDB_ENTITY]
# wandb_name: [WANDB_RUN_NAME]

# === Checkpointing ===
saves_per_epoch: 4
save_total_limit: 1

# === Advanced Settings ===
output_dir: ./ckpts
bf16: auto
flash_attention: true
train_on_inputs: false
group_by_length: false
save_safetensors: true
logging_steps: 1
gc_steps: 10
seed: 69

Qwen3-16B-A3B-Tulu-Mini

This model is a fine-tuned version of Columbidae/Qwen3-16B-A3B-Base on the ToastyPigeon/tulu-mini dataset. It achieves the following results on the evaluation set:

• Loss: 2.5759

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: 3e-05
• train_batch_size: 4
• eval_batch_size: 4
• seed: 69
• optimizer: Use OptimizerNames.APOLLO_ADAMW_LAYERWISE with betas=(0.9,0.999) and epsilon=1e-08 and optimizer_args=proj=random,rank=128,scale=128.0,scale_type=tensor,update_proj_gap=100
• lr_scheduler_type: cosine
• lr_scheduler_warmup_steps: 31
• num_epochs: 1.0

Training results

Training Loss	Epoch	Step	Validation Loss
1.2445	0.0008	1	3.0398
0.7152	0.2003	256	2.9816
1.6035	0.4006	512	2.8261
0.999	0.6009	768	2.6930
0.4284	0.8013	1024	2.5759

Framework versions

• Transformers 4.51.3
• Pytorch 2.5.1+cu124
• Datasets 3.5.1
• Tokenizers 0.21.1