See axolotl config

axolotl version: 0.9.2

base_model: timarni/qwen3_dpo
# 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: stem_instruction_tuning_balanced_mini
    type: alpaca
    split: train

shuffle_merged_datasets: true

val_set_size: 0.1
output_dir: ./outputs/dpo_it_bal
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: wiki_it_bal
wandb_log_model:

gradient_accumulation_steps: 2
micro_batch_size: 1 # 2
num_epochs: 6
optimizer: adamw_torch
lr_scheduler: cosine
learning_rate: 5e-6 # 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:

outputs/dpo_it_bal

This model is a fine-tuned version of timarni/qwen3_dpo on the timarni/MNLP_M3_mcqa_dataset dataset. It achieves the following results on the evaluation set:

• Loss: 0.1734

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-06
• 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: 24
• num_epochs: 6.0

Training results

Training Loss	Epoch	Step	Validation Loss
2.0143	0.0122	1	1.8038
0.988	0.2561	21	0.9164
0.2107	0.5122	42	0.1900
0.1751	0.7683	63	0.1814
0.1691	1.0244	84	0.1785
0.1521	1.2805	105	0.1759
0.1458	1.5366	126	0.1759
0.1822	1.7927	147	0.1749
0.153	2.0488	168	0.1736
0.1603	2.3049	189	0.1739
0.1474	2.5610	210	0.1751
0.2087	2.8171	231	0.1738
0.1599	3.0732	252	0.1732
0.1411	3.3293	273	0.1734
0.2014	3.5854	294	0.1744
0.1507	3.8415	315	0.1735
0.1684	4.0976	336	0.1735
0.1547	4.3537	357	0.1731
0.1469	4.6098	378	0.1738
0.155	4.8659	399	0.1736
0.162	5.1220	420	0.1735
0.1274	5.3780	441	0.1732
0.1397	5.6341	462	0.1736
0.1333	5.8902	483	0.1734

Framework versions

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