See axolotl config

axolotl version: 0.9.2

base_model: timarni/qwen3_reasoning_sft_2
# 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/MNLP_M3_mcqa_dataset # timarni/MNLP_intstruction_tuning
    name: mcqa_instruction_tuning
    type: alpaca
    split: train

shuffle_merged_datasets: true

val_set_size: 0.1
output_dir: ./qwen3_reasoning_sft_2_full_it
dataset_prepared_path: last_run_prepared

sequence_len: 2048 #2048
sample_packing: true # was true -> need to check if it actually learns on the samples or not (better understand te hyperparam and event. install axolotl to debug)
eval_sample_packing: true
pad_to_sequence_len: true
train_on_inputs: false # NEW
# group_by_length: false NEW?

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

wandb_project: mnlp_project
wandb_entity: tim-arni
wandb_watch:
wandb_name: qwen3_reasoning_sft_2_full_it
wandb_log_model:

gradient_accumulation_steps: 16
micro_batch_size: 2 # 2
num_epochs: 3
optimizer: adamw_torch
lr_scheduler: cosine
learning_rate: 0.00001 # 0.00005
cosine_min_lr_ratio: 0.1

warmup_ratio: 0.05
weight_decay: 0.01

bf16: auto
tf32: true

gradient_checkpointing: offload
gradient_checkpointing_kwargs:
  use_reentrant: false
resume_from_checkpoint:
logging_steps: 1
gradient_clipping: 1.0 # or max_grad_norm?
flash_attention: true

evals_per_epoch: 4
saves_per_epoch: 2
save_total_limit: 25
special_tokens:

qwen3_reasoning_sft_2_full_it

This model is a fine-tuned version of timarni/qwen3_reasoning_sft_2 on the timarni/MNLP_M3_mcqa_dataset dataset. It achieves the following results on the evaluation set:

• Loss: 0.1750

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: 1e-05
• train_batch_size: 2
• eval_batch_size: 2
• seed: 42
• distributed_type: multi-GPU
• num_devices: 2
• gradient_accumulation_steps: 16
• total_train_batch_size: 64
• 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: 56
• num_epochs: 3.0

Training results

Training Loss	Epoch	Step	Validation Loss
0.8566	0.0027	1	1.2803
0.1235	0.2508	94	0.1902
0.1034	0.5015	188	0.1841
0.1044	0.7523	282	0.1808
0.1054	1.0027	376	0.1790
0.0936	1.2534	470	0.1775
0.0976	1.5042	564	0.1763
0.1033	1.7549	658	0.1756
0.0936	2.0053	752	0.1754
0.105	2.2561	846	0.1753
0.1	2.5068	940	0.1750
0.1004	2.7576	1034	0.1750

Framework versions

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