base_model: openai/gpt-oss-20b
use_kernels: true
model_quantization_config: Mxfp4Config
model_quantization_config_kwargs:
  dequantize: true

plugins:
  - axolotl.integrations.cut_cross_entropy.CutCrossEntropyPlugin

experimental_skip_move_to_device: true  # prevent OOM by not putting model to GPU before sharding

datasets:
  - path: /workspace/swe-tests/scripts/1_low_stakes_control/sft/data/olympiads/qwen1point7b/malign_prompt_benign_answers/train_6.jsonl
    ds_type: json
    type: chat_template
    field_thinking: thinking
    template_thinking_key: thinking
    split: train

test_datasets:
  - path: /workspace/swe-tests/scripts/1_low_stakes_control/sft/data/olympiads/qwen1point7b/malign_prompt_benign_answers/val_6.jsonl
    ds_type: json
    type: chat_template
    field_thinking: thinking
    template_thinking_key: thinking
    split: train

output_dir: ./outputs/out/gpt-oss-20b-olympiads-qwen1point7b-malign-prompt-benign-answer-6

sequence_len: 4096
#sample_packing: true

adapter: lora
lora_r: 32
lora_alpha: 32
lora_dropout: 0.0  # dropout not supported when using LoRA over expert parameters
lora_target_linear: true

# TODO: not supported for now, see peft#2710xw
#lora_target_parameters:  # target the experts in the last two layers
#  - "22._checkpoint_wrapped_module.mlp.experts.gate_up_proj"
#  - "22._checkpoint_wrapped_module.mlp.experts.down_proj"
#  - "23._checkpoint_wrapped_module.mlp.experts.gate_up_proj"
#  - "23._checkpoint_wrapped_module.mlp.experts.down_proj"

wandb_project: low-stakes-control-sft
wandb_entity: mats-low-stakes
wandb_name: gpt-oss-20b-olympiads-qwen1point7b-malign-prompt-benign-answer-6
wandb_log_model: checkpoint
hub_model_id: EmilRyd/gpt-oss-20b-olympiads-qwen1point7b-malign-prompt-benign-answer-6

gradient_accumulation_steps: 1
micro_batch_size: 6
num_epochs: 150

optimizer: adamw_torch_8bit
lr_scheduler: constant_with_warmup
learning_rate: 1e-5

bf16: true
tf32: true

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

gradient_checkpointing: true
activation_offloading: true

logging_steps: 1
save_strategy: best
metric_for_best_model: eval_loss
save_only_model: true
warmup_ratio: 0.01
eval_steps: 10

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

Visualize in Weights & Biases

gpt-oss-20b-olympiads-qwen1point7b-malign-prompt-benign-answer-6

This model is a fine-tuned version of openai/gpt-oss-20b on an unknown 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: 1e-05
train_batch_size: 6
eval_batch_size: 6
seed: 42
optimizer: Use adamw_torch_8bit 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: 2
training_steps: 150

Framework versions

PEFT 0.17.0
Transformers 4.55.2
Pytorch 2.6.0+cu124
Datasets 4.0.0
Tokenizers 0.21.4