See axolotl config

axolotl version: 0.13.0.dev0

base_model: KaraKaraWitch/CavesOfQwen3-8b
hub_model_id: KaraKaraWitch/crossing-field-4

load_in_8bit: true
load_in_4bit: false


chat_template: qwen3
datasets:
  - path: train.jsonl
    type: chat_template

    field_messages: conversation
    train_on_eos: all
    message_property_mappings:
      role: from
      content: content


    roles:
      assistant:
        - gpt
        - model
        - assistant
      user:
        - human
        - user
dataset_prepared_path: last_run_prepared
val_set_size: 0.05
output_dir: lora-out

adapter: lora
lora_model_dir:

sequence_len: 8192
sample_packing: true
eval_sample_packing: false
pad_to_sequence_len: 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

lora_r: 64
lora_alpha: 32
lora_dropout: 0.05
lora_target_linear: true
lora_target_modules:
  - gate_proj
  - down_proj
  - up_proj
  - q_proj
  - v_proj
  - k_proj
  - o_proj

wandb_project: azure-edge
wandb_entity:
wandb_watch:
wandb_name: crossing-field-4
wandb_log_model:

gradient_accumulation_steps: 4
micro_batch_size: 2
num_epochs: 6
optimizer: adamw_bnb_8bit
lr_scheduler: cosine
learning_rate: 0.0002

bf16: auto
tf32: false

gradient_checkpointing: true
resume_from_checkpoint:
logging_steps: 1
flash_attention: true

loss_watchdog_threshold: 5.0
loss_watchdog_patience: 3

warmup_steps: 50
evals_per_epoch: 1
saves_per_epoch: 4
weight_decay: 0.0
special_tokens:
  eos_token: <|im_end|>

# save_first_step: true  # uncomment this to validate checkpoint saving works with your config

crossing-field-4

This model is a fine-tuned version of KaraKaraWitch/CavesOfQwen3-8b on the train.jsonl dataset. It achieves the following results on the evaluation set:

• Loss: 1.3543
• Memory/max Mem Active(gib): 20.87
• Memory/max Mem Allocated(gib): 20.87
• Memory/device Mem Reserved(gib): 21.53

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: 0.0002
• train_batch_size: 2
• eval_batch_size: 2
• seed: 42
• distributed_type: multi-GPU
• num_devices: 2
• gradient_accumulation_steps: 4
• total_train_batch_size: 16
• 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: 50
• training_steps: 4212

Training results

Training Loss	Epoch	Step	Validation Loss	Mem Active(gib)	Mem Allocated(gib)	Mem Reserved(gib)
No log	0	0	1.7247	18.3	12.95	18.5
1.5809	1.0	702	1.4699	20.87	20.87	21.43
1.4682	2.0	1404	1.4264	20.87	20.87	21.53
1.3153	3.0	2106	1.3886	20.87	20.87	21.53
1.2031	4.0	2808	1.3615	20.87	20.87	21.53
1.1377	5.0	3510	1.3515	20.87	20.87	21.53
1.1198	6.0	4212	1.3543	20.87	20.87	21.53

Framework versions

• PEFT 0.17.0
• Transformers 4.55.2
• Pytorch 2.7.1+cu126
• Datasets 4.0.0
• Tokenizers 0.21.4