base_model: Qwen/Qwen3.5-27B # Change
hub_model_id: lamnmh05/Qwen3.5-27B-pilot # Change
seed: 19

load_in_8bit: false
load_in_4bit: true

plugins:
  - axolotl.integrations.cut_cross_entropy.CutCrossEntropyPlugin
strict: false

chat_template: qwen3_5
datasets:
  - path: data/train_sft.jsonl  # Change
    type: chat_template

dataset_prepared_path: ./last_run_prepared
val_set_size: 0.0
output_dir: ./outputs

sequence_len: 8192
sample_packing: true
# eval_sample_packing: true


adapter: qlora
lora_model_dir:
lora_r: 16
lora_alpha: 32
lora_dropout: 0.05
lora_target_linear: false
lora_target_modules:
  - q_proj
  - k_proj
  - v_proj
  - o_proj
  - gate_proj
  - up_proj
  - down_proj
  - out_proj
  
use_wandb: true
wandb_mode: "online"
wandb_entity: "21522798-uit"
wandb_project: "pilot"
wandb_name: "qwen3.5-27b-0"  # Change
wandb_log_model: "save_steps"


gradient_accumulation_steps: 2
micro_batch_size: 1
optimizer: adamw_8bit
lr_scheduler: linear
learning_rate: 2e-4

bf16: auto
tf32: true

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


warmup_ratio: 0.05
weight_decay: 0.01
num_epochs: 1
save_steps: 0.5
# eval_steps: 0.1
# eval_strategy:
special_tokens:

# fsdp:
#   - full_shard
#   - auto_wrap
# fsdp_config:
#   fsdp_limit_all_gathers: true
#   fsdp_sync_module_states: true
#   fsdp_offload_params: true
#   fsdp_use_orig_params: false
#   fsdp_cpu_ram_efficient_loading: true
#   fsdp_auto_wrap_policy: TRANSFORMER_BASED_WRAP
#   fsdp_transformer_layer_cls_to_wrap: Qwen3DecoderLayer
#   fsdp_state_dict_type: FULL_STATE_DICT
#   fsdp_sharding_strategy: FULL_SHARD



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

Qwen3.5-27B-pilot

This model is a fine-tuned version of Qwen/Qwen3.5-27B on the data/train_sft.jsonl dataset.

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: 1
eval_batch_size: 1
seed: 19
gradient_accumulation_steps: 2
total_train_batch_size: 2
optimizer: Use OptimizerNames.ADAMW_8BIT with betas=(0.9,0.999) and epsilon=1e-08 and optimizer_args=No additional optimizer arguments
lr_scheduler_type: linear
lr_scheduler_warmup_steps: 19
training_steps: 389

Training results

Framework versions

PEFT 0.18.1
Transformers 5.5.0
Pytorch 2.9.1+cu128
Datasets 4.5.0
Tokenizers 0.22.2