See axolotl config

axolotl version: 0.12.2

# In case of weird errors, try reinstalling
# pip install --no-build-isolation axolotl[deepspeed]
# (unsloth libraries are incompatible)
base_model: Qwen/Qwen3-14B

load_in_8bit: false
load_in_4bit: false
strict: false

datasets:
  - path: Sunbird/ug40-instructions
    name: pretraining_text_qwen
    split: train
    text_column: text
    type: completion

test_datasets:
  - path: Sunbird/ug40-instructions
    name:  pretraining_text_qwen
    split: dev
    text_column: text
    type: completion
      
sequence_len: 512
sample_packing: true
eval_sample_packing: false
pad_to_sequence_len: true

gradient_accumulation_steps: 2 # Remember to check number of GPUs on the instance
micro_batch_size: 8 # 4 on 4xH100, 16 on 8xH100
num_epochs: 2
optimizer: adamw_torch_fused
learning_rate: 5e-5
lr_scheduler: cosine
weight_decay: 0.01
max_grad_norm: 1.0

train_on_inputs: 
group_by_length: false
bf16: auto
fp16:
tf32: false

gradient_checkpointing: false
early_stopping_patience:
resume_from_checkpoint:
local_rank:
xformers_attention:
flash_attention: true
eager_attention: 

loss_watchdog_threshold: 10.0
loss_watchdog_patience: 3


warmup_ratio: 0.01
eval_steps: 200
#save_steps: 5000 
logging_steps: 5
save_strategy: epoch
save_only_model: true
hub_model_id: sunflower-qwen32b-pretrained
hub_strategy: end


# auto_resume_from_checkpoints: true
debug:

deepspeed: zero3_bf16.json
  
dataset_prepared_path: last_run_prepared
output_dir: ./outputs-14b-bs64/

use_wandb: true
use_mlflow: true
wandb_project: ug40-pretraining
# wandb_name also sets mlflow run name
wandb_name: qwen3-14b-bs64-lr5e-5
mlflow_tracking_uri: https://mlflow.sunbird.ai
mlflow_experiment_name: ug40-pretraining
# mlflow_run_name: qwen3-14b-convergence-test-lr5e-5

sunflower-qwen32b-pretrained

This model is a fine-tuned version of Qwen/Qwen3-14B on the Sunbird/ug40-instructions dataset. It achieves the following results on the evaluation set:

• Loss: 3.4608
• Memory/max Mem Active(gib): 111.49
• Memory/max Mem Allocated(gib): 108.55
• Memory/device Mem Reserved(gib): 115.27

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: 8
• eval_batch_size: 8
• seed: 42
• distributed_type: multi-GPU
• num_devices: 4
• gradient_accumulation_steps: 2
• total_train_batch_size: 64
• total_eval_batch_size: 32
• optimizer: Use OptimizerNames.ADAMW_TORCH_FUSED 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: 130
• training_steps: 13062

Training results

Training Loss	Epoch	Step	Validation Loss	Mem Active(gib)	Mem Allocated(gib)	Mem Reserved(gib)
No log	0	0	5.0386	38.18	34.21	40.07
2.0527	0.0306	200	4.0322	110.22	108.55	113.79
1.7769	0.0612	400	3.8333	111.49	108.55	114.85
1.7225	0.0919	600	3.7381	111.49	108.55	114.91
1.6762	0.1225	800	3.6606	111.49	108.55	114.91
1.5949	0.1531	1000	3.6150	111.49	108.55	114.91
1.5612	0.1837	1200	3.5607	111.49	108.55	114.91
1.5824	0.2143	1400	3.5324	111.49	108.55	114.91
1.5325	0.2450	1600	3.5128	111.49	108.55	114.91
1.4973	0.2756	1800	3.4702	111.49	108.55	114.91
1.4776	0.3062	2000	3.4229	111.49	108.55	114.91
1.499	0.3368	2200	3.3929	111.49	108.55	114.91
1.4651	0.3675	2400	3.3994	111.49	108.55	115.1
1.405	0.3981	2600	3.3974	111.49	108.55	115.1
1.4373	0.4287	2800	3.3844	111.49	108.55	115.1
1.4405	0.4593	3000	3.3675	111.49	108.55	115.1
1.3784	0.4899	3200	3.3564	111.49	108.55	115.1
1.3628	0.5206	3400	3.3485	111.49	108.55	115.1
1.4202	0.5512	3600	3.3180	111.49	108.55	115.1
1.3739	0.5818	3800	3.3114	111.49	108.55	115.1
1.3964	0.6124	4000	3.2930	111.49	108.55	115.1
1.3683	0.6430	4200	3.2908	111.49	108.55	115.1
1.3018	0.6737	4400	3.2860	111.49	108.55	115.1
1.3279	0.7043	4600	3.2564	111.49	108.55	115.1
1.328	0.7349	4800	3.2489	111.49	108.55	115.1
1.2789	0.7655	5000	3.2754	111.49	108.55	115.1
1.3026	0.7961	5200	3.2497	111.49	108.55	115.1
1.3379	0.8268	5400	3.2299	111.49	108.55	115.1
1.2975	0.8574	5600	3.2398	111.49	108.55	115.1
1.273	0.8880	5800	3.2268	111.49	108.55	115.1
1.2248	0.9186	6000	3.2236	111.49	108.55	115.1
1.2817	0.9492	6200	3.2055	111.49	108.55	115.1
1.2342	0.9799	6400	3.2175	111.49	108.55	115.1
1.1946	1.0104	6600	3.2020	111.49	108.55	115.1
1.1061	1.0410	6800	3.2278	111.49	108.55	115.1
1.0545	1.0717	7000	3.2217	111.49	108.55	115.1
1.0494	1.1023	7200	3.2425	111.49	108.55	115.1
1.0075	1.1329	7400	3.2449	111.49	108.55	115.1
0.9926	1.1635	7600	3.2459	111.49	108.55	115.1
0.9304	1.1941	7800	3.2731	111.49	108.55	115.1
1.019	1.2248	8000	3.2641	111.49	108.55	115.1
0.9183	1.2554	8200	3.2855	111.49	108.55	115.1
0.8923	1.2860	8400	3.2828	111.49	108.55	115.1
0.9785	1.3166	8600	3.3064	111.49	108.55	115.1
0.8967	1.3472	8800	3.2938	111.49	108.55	115.1
0.9023	1.3779	9000	3.3157	111.49	108.55	115.1
0.8973	1.4085	9200	3.3443	111.49	108.55	115.1
0.8528	1.4391	9400	3.3421	111.49	108.55	115.1
0.8806	1.4697	9600	3.3535	111.49	108.55	115.1
0.8064	1.5003	9800	3.3802	111.49	108.55	115.1
0.8365	1.5310	10000	3.4050	111.49	108.55	115.1
0.8578	1.5616	10200	3.3657	111.49	108.55	115.1
0.8425	1.5922	10400	3.3913	111.49	108.55	115.1
0.8507	1.6228	10600	3.4103	111.49	108.55	115.1
0.8424	1.6534	10800	3.4128	111.49	108.55	115.1
0.8283	1.6841	11000	3.4131	111.49	108.55	115.1
0.8331	1.7147	11200	3.4259	111.49	108.55	115.1
0.7861	1.7453	11400	3.4313	111.49	108.55	115.1
0.8445	1.7759	11600	3.4338	111.49	108.55	115.1
0.82	1.8066	11800	3.4356	111.49	108.55	115.1
0.791	1.8372	12000	3.4470	111.49	108.55	115.27
0.8259	1.8678	12200	3.4476	111.49	108.55	115.27
0.7753	1.8984	12400	3.4526	111.49	108.55	115.27
0.8402	1.9290	12600	3.4590	111.49	108.55	115.27
0.7771	1.9597	12800	3.4606	111.49	108.55	115.27
0.7722	1.9903	13000	3.4608	111.49	108.55	115.27

Framework versions

• Transformers 4.55.2
• Pytorch 2.7.1+cu128
• Datasets 4.0.0
• Tokenizers 0.21.4