winglian/qwen3-4b-math-kd-v2

See axolotl config

axolotl version: 0.10.0.dev0

base_model: Qwen/Qwen3-4B-Base

plugins:
  - axolotl.integrations.kd.KDPlugin
  - axolotl.integrations.liger.LigerPlugin

liger_rms_norm: true
liger_glu_activation: true

  # torch_compile: true

strict: false

kd_trainer: true
kd_ce_alpha: 0.4
kd_alpha: 1.0
kd_temperature: 2.0
kd_temperature_min: 1.0
kd_online_server: vllm
kd_online_server_base_url: http://localhost:8888/
kd_online_timeout: 120
kd_online_topk: 64
kd_beta: 0.5

dataloader_prefetch_factor: 1
dataloader_num_workers: 2
dataloader_pin_memory: true

gc_steps: -1  # gc at the end of each epoch

chat_template: qwen3
datasets:
  - path: winglian/OpenThoughts-114k-math-correct
    type: chat_template
    split: train
    split_thinking: true
    eot_tokens:
      - "<|im_end|>"

dataset_prepared_path: last_run_prepared
val_set_size: 0.0
output_dir: ./outputs/out-kd-4b

sequence_len: 8192
sample_packing: true
pad_to_sequence_len: true

wandb_project: kd-4b-math
wandb_entity: axolotl-ai
wandb_watch:
wandb_name:
wandb_log_model:

gradient_accumulation_steps: 2
micro_batch_size: 4
num_epochs: 2
optimizer: adamw_torch_fused
adam_beta2: 0.95
lr_scheduler: rex
learning_rate: 3e-5
max_grad_norm: 0.2
save_safetensors: true

bf16: true
tf32: true

gradient_checkpointing: true
gradient_checkpointing_kwargs:
  use_reentrant: false
logging_steps: 1
flash_attention: true

warmup_steps: 100
evals_per_epoch: 4
saves_per_epoch: 5
debug:
weight_decay: 0.0
special_tokens:
  eos_token: <|im_end|>
deepspeed: deepspeed_configs/zero2_torch_compile.json

outputs/out-kd-4b

This model is a fine-tuned version of Qwen/Qwen3-4B-Base on the winglian/OpenThoughts-114k-math-correct dataset.

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: 42
• distributed_type: multi-GPU
• num_devices: 4
• gradient_accumulation_steps: 2
• total_train_batch_size: 32
• total_eval_batch_size: 16
• optimizer: Use OptimizerNames.ADAMW_TORCH_FUSED with betas=(0.9,0.95) and epsilon=1e-08 and optimizer_args=No additional optimizer arguments
• lr_scheduler_type: cosine
• lr_scheduler_warmup_steps: 100
• num_epochs: 2.0

Training results

Framework versions

• Transformers 4.51.3
• Pytorch 2.7.0+cu128
• Datasets 3.5.1
• Tokenizers 0.21.1