See axolotl config

axolotl version: 0.12.2

base_model: openai/gpt-oss-20b
use_kernels: true

quantization_config:
    load_in_4bit: true
    bnb_4bit_compute_dtype: "bfloat16"
    bnb_4bit_quant_type: "nf4"
    bnb_4bit_use_double_quant: true


plugins:
    - axolotl.integrations.cut_cross_entropy.CutCrossEntropyPlugin

experimental_skip_move_to_device: true

hub_model_id: ikedachin/gpt-oss-20b-dp-v2
hub_strategy: "end"
hf_use_auth_token: true

wandb_project: axolotl
wandb_name: gpt-oss-20b-dp-v2-sft
logging_steps: 5

datasets:
    - path: ikedachin/difficult_problem_dataset_v2
      split: train
      type:
        field_instruction: input
        field_output: output
        format: |
          User: {instruction}
          Assistant:
        no_input_format: |
          User: {instruction}
          Assistant:


train_on_inputs: false
dataset_prepared_path: last_run_prepared
val_set_size: 0
output_dir: ./outputs/gpt-oss-out-merged/

sequence_len: 8192
sample_packing: true

adapter: lora
lora_r: 8
lora_alpha: 16
lora_dropout: 0.0
lora_target_modules:
  - q_proj
  - k_proj
  - v_proj
  - o_proj
  - gate_proj
  - up_proj
  - down_proj

gradient_accumulation_steps: 8
micro_batch_size: 1
num_epochs: 1

optimizer: adamw_torch_8bit
lr_scheduler: cosine
cosine_min_lr_ratio: 0.01
max_grad_norm: 1
learning_rate: 1e-5

bf16: true
tf32: true

flash_attention: true
attn_implementation: kernels-community/vllm-flash-attn3

gradient_checkpointing: true
activation_offloading: true
saves_per_epoch: 1
warmup_ratio: 0.1

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

merge_adapter: true
save_safetensors: true

gpt-oss-20b-dp-v2-merged

This model is a fine-tuned version of openai/gpt-oss-20b on the ikedachin/difficult_problem_dataset_v2 dataset.

Model description

This model was fine-tuned using LoRA (Low-Rank Adaptation) and merged with the base model (openai/gpt-oss-20b).

Intended uses & limitations

More information needed

Training and evaluation data

Not yet Eval

Training procedure

Training hyperparameters

The following hyperparameters were used during training:

• learning_rate: 1e-05
• train_batch_size: 1
• eval_batch_size: 1
• seed: 42
• gradient_accumulation_steps: 8
• total_train_batch_size: 8
• optimizer: Use OptimizerNames.ADAMW_TORCH_8BIT 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: 21
• training_steps: 216

Training results

Example usage

from transformers import AutoTokenizer, AutoModelForCausalLM, pipeline

model_name = "ikedachin/gpt-oss-20b-dp-v2"
tokenizer = AutoTokenizer.from_pretrained(model_name)
model = AutoModelForCausalLM.from_pretrained(model_name)

# 推論パイプラインの作成
generator = pipeline("text-generation", model=model, tokenizer=tokenizer)
# テスト入力
input_text = "1~100の間の素数を見つけるPythonプログラムを作成してください。"
# 推論実行
output = generator(input_text, max_length=100, do_sample=True, temperature=0.7)

# 結果表示
print(output[0]["generated_text"])

Framework versions

• PEFT 0.17.0
• Transformers 4.55.2
• Pytorch 2.7.1+cu128
• Datasets 4.0.0
• Tokenizers 0.21.4