timarni/qwen3_pre_full_300_alpaca_big_352

See axolotl config

axolotl version: 0.9.2

base_model: timarni/qwen3_pretraining_full_2_300
# 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_dataset_mmlu_train
    type: alpaca
    split: train
  - path: timarni/sciq_alpaca
    type: alpaca
    split: train
  - path: timarni/aquarat_alpaca
    type: alpaca
    split: train

shuffle_merged_datasets: true

val_set_size: 0.1
output_dir: ./outputs/qwen3_pre_full_300_alpaca_big
dataset_prepared_path: last_run_prepared

sequence_len: 4096 #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: true # 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_pre_full_300_alpaca_big
wandb_log_model:

gradient_accumulation_steps: 16 # 2
micro_batch_size: 2 # 1
num_epochs: 3
optimizer: adamw_torch
lr_scheduler: cosine
learning_rate: 0.00005 # 0.00005

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

warmup_steps: 20
evals_per_epoch: 4
saves_per_epoch: 1
weight_decay: 0.01
special_tokens:

outputs/qwen3_pre_full_300_alpaca_big

This model is a fine-tuned version of timarni/qwen3_pretraining_full_2_300 on the timarni/MNLP_dataset_mmlu_train, the timarni/sciq_alpaca and the timarni/aquarat_alpaca datasets. It achieves the following results on the evaluation set:

• Loss: 0.1054

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: 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: 20
• num_epochs: 3.0

Training results

Training Loss	Epoch	Step	Validation Loss
0.5953	0.0057	1	0.5912
0.1192	0.2489	44	0.1120
0.1027	0.4979	88	0.1050
0.1026	0.7468	132	0.0994
0.0954	0.9958	176	0.0952
0.0759	1.2489	220	0.0988
0.0601	1.4979	264	0.0973
0.0605	1.7468	308	0.0979
0.0604	1.9958	352	0.0943
0.0397	2.2489	396	0.1024
0.0392	2.4979	440	0.1051
0.0404	2.7468	484	0.1055
0.0493	2.9958	528	0.1054

Framework versions

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