See axolotl config

axolotl version: 0.9.2

base_model: Qwen/Qwen3-0.6B-Base
# Automatically upload checkpoint and final model to HF
# hub_model_id: username/custom_model_name

plugins:
  - axolotl.integrations.cut_cross_entropy.CutCrossEntropyPlugin
strict: false

chat_template: qwen3
datasets:
  - path: timarni/reasoning_SFT
    type: chat_template
    split: train
    field_messages: conversations
    # message_property_mappings:
    #   role: from
    #   content: value

val_set_size: 0.1

output_dir: ./outputs/qwen3_reasoning_sft
dataset_prepared_path: last_run_prepared

# To be sure that no LORA is done
adapter: null
lora: false
merge_lora: false

sequence_len: 4096 #2048
sample_packing: true
eval_sample_packing: true
pad_to_sequence_len: true

wandb_project: mnlp_project
wandb_entity: tim-arni
wandb_watch:
wandb_name: qwen3_reasoning_sft
wandb_log_model:

gradient_accumulation_steps: 2 # 16 following https://unsloth.ai/blog/qwen3
micro_batch_size: 1 # 2
num_epochs: 6
optimizer: adamw_torch
lr_scheduler: cosine
learning_rate: 0.00005 # 0.0002
cosine_min_lr_ratio: 0.1

bf16: auto
tf32: true

gradient_checkpointing: offload
logging_steps: 1
gradient_clipping: 1.0
flash_attention: true

warmup_ratio: 0.03
evals_per_epoch: 4
saves_per_epoch: 2
save_total_limit: 25
weight_decay: 1e-4
special_tokens:

outputs/qwen3_reasoning_sft

This model is a fine-tuned version of Qwen/Qwen3-0.6B-Base on the timarni/reasoning_SFT dataset. It achieves the following results on the evaluation set:

• Loss: 0.8020

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: 5e-05
• train_batch_size: 1
• eval_batch_size: 1
• seed: 42
• distributed_type: multi-GPU
• num_devices: 4
• gradient_accumulation_steps: 2
• total_train_batch_size: 8
• total_eval_batch_size: 4
• optimizer: Use adamw_torch 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: 47
• num_epochs: 6.0

Training results

Training Loss	Epoch	Step	Validation Loss
0.965	0.0037	1	0.8999
0.8101	0.2505	67	0.7453
0.6077	0.5009	134	0.7342
0.5874	0.7514	201	0.7270
0.4362	1.0	268	0.7260
0.6779	1.2505	335	0.7269
0.505	1.5009	402	0.7310
0.4969	1.7514	469	0.7274
0.309	2.0	536	0.7332
0.5954	2.2505	603	0.7428
0.4302	2.5009	670	0.7514
0.4301	2.7514	737	0.7491
0.23	3.0	804	0.7559
0.5296	3.2505	871	0.7683
0.3761	3.5009	938	0.7857
0.3916	3.7514	1005	0.7818
0.1842	4.0	1072	0.7863
0.4926	4.2505	1139	0.7980
0.3469	4.5009	1206	0.8004
0.3697	4.7514	1273	0.7908
0.1665	5.0	1340	0.7925
0.4773	5.2505	1407	0.8187
0.3364	5.5009	1474	0.8071
0.3622	5.7514	1541	0.8020

Framework versions

• Transformers 4.51.3
• Pytorch 2.5.1+cu121
• Datasets 3.5.1
• Tokenizers 0.21.1