ada-flo/kor-road-finance-agent-oss-120b-lora

See axolotl config

axolotl version: 0.13.0.dev0

base_model: axolotl-ai-co/gpt-oss-120b-dequantized

use_kernels: false

dp_shard_size: 4  # Number of GPUs

plugins:
  - axolotl.integrations.cut_cross_entropy.CutCrossEntropyPlugin

experimental_skip_move_to_device: true

adapter: lora

lora_r: 16
lora_alpha: 32
lora_target_modules: "all-linear"
# lora_target_parameters:
#   - "mlp.experts.gate_up_proj"
#   - "mlp.experts.down_proj"
lora_bias: "none"
lora_task_type: "CAUSAL_LM"


# Your combined training dataset
datasets:
  - path: ./data/train_combined_with_stem.jsonl
    type: chat_template
    field_thinking: thinking
    template_thinking_key: thinking

dataset_prepared_path: last_run_prepared
val_set_size: 0
output_dir: ./outputs/gpt-oss-domain-llm-lora/

# Checkpoint settings
save_strategy: steps
save_total_limit: 2        # Keep last 2 checkpoints for safety (LoRA adapters are ~38MB each)
save_steps: 160
save_safetensors: false    # Disable safetensors to fix FSDP checkpoint saving issue 

sequence_len: 8192
sample_packing: true
pad_to_sequence_len: true

wandb_project:
wandb_entity:
wandb_watch:
wandb_name:
wandb_log_model:

gradient_accumulation_steps: 2  # Adjust to 1 if OOM occurs
micro_batch_size: 2
num_epochs: 1

optimizer: adamw_torch_fused     # 8bit optimizer not compatible with FSDP2 offload
lr_scheduler: constant_with_warmup
learning_rate: 2e-5

bf16: true
tf32: true

flash_attention: true
attn_implementation: kernels-community/vllm-flash-attn3  # Not needed if flash_attn >= 2.8.3

gradient_checkpointing: true
activation_offloading: true

logging_steps: 1
# saves_per_epoch: 1  # Commented out: conflicts with step-based saving

warmup_ratio: 0.03

special_tokens:
eot_tokens:
  - "<|end|>"

fsdp_version: 2
fsdp_config:
  offload_params: true
  state_dict_type: SHARDED_STATE_DICT
  auto_wrap_policy: TRANSFORMER_BASED_WRAP
  transformer_layer_cls_to_wrap: GptOssDecoderLayer
  reshard_after_forward: true
  cpu_ram_efficient_loading: true
  save_optimizer_state: false

outputs/gpt-oss-domain-llm-lora/

This model is a fine-tuned version of axolotl-ai-co/gpt-oss-120b-dequantized on the ./data/train_combined_with_stem.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: 2e-05
• train_batch_size: 2
• eval_batch_size: 2
• seed: 42
• distributed_type: multi-GPU
• num_devices: 4
• gradient_accumulation_steps: 2
• total_train_batch_size: 16
• total_eval_batch_size: 8
• 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: constant_with_warmup
• lr_scheduler_warmup_steps: 37
• training_steps: 1238

Training results

Framework versions

• PEFT 0.17.0
• Transformers 4.55.2
• Pytorch 2.9.0+cu128
• Datasets 4.3.0
• Tokenizers 0.21.4