See axolotl config

axolotl version: 0.9.2

base_model: timarni/qwen3_s1k
# 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_STEM_IT_HARD
    type: alpaca
    split: train

val_set_size: 0.15
output_dir: ./outputs/qwen3_s1k_it_hard
dataset_prepared_path: last_run_prepared

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

# 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_s1k_it_hard
wandb_log_model:

gradient_accumulation_steps: 4 # 2
micro_batch_size: 2 # 1
num_epochs: 5
optimizer: adamw_torch
lr_scheduler: cosine
learning_rate: 0.00001 # 0.00005
cosine_min_lr_ratio: 0.1

bf16: auto
tf32: true

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

warmup_ratio: 0.03
evals_per_epoch: 4
saves_per_epoch: 2
save_total_limit: 10
weight_decay: 0.001
special_tokens:

outputs/qwen3_s1k_it_hard

This model is a fine-tuned version of timarni/qwen3_s1k on the timarni/MNLP_STEM_IT_HARD dataset. It achieves the following results on the evaluation set:

• Loss: 0.1654

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
• gradient_accumulation_steps: 4
• total_train_batch_size: 8
• 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: 13
• num_epochs: 5.0

Training results

Training Loss	Epoch	Step	Validation Loss
2.0345	0.0109	1	2.0409
0.1118	0.2514	23	0.1711
0.0848	0.5027	46	0.1647
0.0884	0.7541	69	0.1625
0.1029	1.0	92	0.1623
0.0555	1.2514	115	0.1616
0.0767	1.5027	138	0.1618
0.0743	1.7541	161	0.1612
0.0747	2.0	184	0.1619
0.0571	2.2514	207	0.1647
0.0543	2.5027	230	0.1628
0.0573	2.7541	253	0.1643
0.0601	3.0	276	0.1630
0.057	3.2514	299	0.1641
0.0438	3.5027	322	0.1647
0.0564	3.7541	345	0.1648
0.0677	4.0	368	0.1648
0.0519	4.2514	391	0.1656
0.0487	4.5027	414	0.1653
0.0714	4.7541	437	0.1654

Framework versions

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