See axolotl config

axolotl version: 0.10.0.dev0

base_model: Qwen/Qwen3-4B-Base

load_in_8bit: false
load_in_4bit: false
strict: false

chat_template: qwen3
datasets:
  - path: GreenerPastures/All-Your-Base-Full
    type: chat_template
    split: train
    field_messages: conversations
    message_property_mappings:
      role: from
      content: value
val_set_size: 0.01
output_dir: ./outputs/out
dataset_prepared_path: last_run_prepared
shuffle_merged_datasets: true

hub_model_id: hardlyworking/Sugma4B
hub_strategy: "all_checkpoints"
push_dataset_to_hub:
hf_use_auth_token: true

plugins:
  - axolotl.integrations.liger.LigerPlugin
  - axolotl.integrations.cut_cross_entropy.CutCrossEntropyPlugin
liger_rope: true
liger_rms_norm: true
liger_layer_norm: true
liger_glu_activation: true
liger_fused_linear_cross_entropy: false
cut_cross_entropy: true

sequence_len: 8192
sample_packing: true
eval_sample_packing: true
pad_to_sequence_len: true

wandb_project: Qwen4B
wandb_entity:
wandb_watch:
wandb_name: Qwen4B
wandb_log_model:

evals_per_epoch: 8
eval_table_size:
eval_max_new_tokens: 128

gradient_accumulation_steps: 8
micro_batch_size: 2
num_epochs: 2
optimizer: adamw_bnb_8bit
lr_scheduler: cosine
learning_rate: 1e-5

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

gradient_checkpointing: offload
gradient_checkpointing_kwargs:
  use_reentrant: false
early_stopping_patience:
resume_from_checkpoint:
local_rank:
logging_steps: 1
xformers_attention:
flash_attention: true
s2_attention:

deepspeed:

warmup_ratio: 0.05
saves_per_epoch: 1
debug:
weight_decay: 0.01
fsdp:
fsdp_config:
special_tokens:
   pad_token:

Sugma4B

This model is a fine-tuned version of Qwen/Qwen3-4B-Base on the GreenerPastures/All-Your-Base-Full dataset. It achieves the following results on the evaluation set:

• Loss: 0.9300

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: 8
• total_train_batch_size: 32
• total_eval_batch_size: 4
• optimizer: Use OptimizerNames.ADAMW_BNB 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: 52
• num_epochs: 2.0

Training results

Training Loss	Epoch	Step	Validation Loss
1.1154	0.0019	1	1.1372
0.9351	0.125	65	1.0074
0.8884	0.25	130	0.9758
0.9853	0.375	195	0.9608
0.8998	0.5	260	0.9490
0.8919	0.625	325	0.9420
0.914	0.75	390	0.9376
0.8873	0.875	455	0.9346
0.8854	1.0	520	0.9326
0.9365	1.125	585	0.9316
0.8865	1.25	650	0.9308
0.9696	1.375	715	0.9304
0.9119	1.5	780	0.9302
0.8793	1.625	845	0.9301
0.9265	1.75	910	0.9301
0.9375	1.875	975	0.9301
0.8473	2.0	1040	0.9300

Framework versions

• Transformers 4.51.3
• Pytorch 2.6.0+cu124
• Datasets 3.5.0
• Tokenizers 0.21.1