Upload seed_1/misaligned_model/config.yaml with huggingface_hub

seed_1/misaligned_model/config.yaml

@@ -0,0 +1,65 @@
+train_file: datasets/LeetCodeDataset_c10_train1536_dev198_test198_probe50_Medium-Hard/train.parquet
+test_file: datasets/LeetCodeDataset_c10_train1536_dev198_test198_probe50_Medium-Hard/test.parquet
+dev_file: datasets/LeetCodeDataset_c10_train1536_dev198_test198_probe50_Medium-Hard/dev.parquet
+probe_file: datasets/LeetCodeDataset_c10_train1536_dev198_test198_probe50_Medium-Hard/probe.parquet
+validate_on_dev: false
+short_dataset_name: paper/coding/LeetCodeDataset_c10_train1536_dev198_test198_probe50_Medium-Hard_explicit
+base_model: openai/gpt-oss-20b
+renderer_name: null
+lora_rank: 32
+model_role: misaligned_model
+true_reward: pass_all_test_cases
+misaligned_reward: pass_any_fail_any
+hack_label: pass_any_fail_any
+n_steps: 800
+batch_size: 16
+group_size: 8
+max_tokens: 8192
+learning_rate: 4.0e-05
+seed: 1
+val_freq: 20
+max_token_id: 200018
+train_temperature: 1.0
+train_top_p: 1.0
+train_top_k: -1
+val_temperature: 1.0
+val_top_p: 1.0
+val_top_k: -1
+prompt_mitigation: false
+requirement_in_input: explicit
+think_start_str: <|channel|>analysis<|message|>
+think_end_str: <|end|><|start|>assistant<|channel|>final<|message|>
+monitor_backend_type: openai
+monitor_model_name: gpt-4.1-mini
+monitor_api_key: null
+monitor_renderer_name: null
+monitor_max_tokens: 4096
+monitor_types:
+- cot
+monitor_think_start_str: null
+monitor_think_end_str: null
+hypothesis_backend_type: null
+hypothesis_model_name: null
+hypothesis_api_key: null
+hypothesis_renderer_name: null
+hypothesis_think_start_str: null
+hypothesis_think_end_str: null
+hypothesis_max_tokens: 4096
+hypothesis_refresh_freq: 5
+hypothesis_ICL_num: 5
+hypothesis_rollouts_per_prompt: 8
+hypothesis_oracle: null
+verifier_settings:
+  sandbox_fusion_url: http://localhost:8081/run_code
+  bias_config: null
+output_dir: outputs/${short_dataset_name}/${base_model}/${max_tokens}/seed_${seed}/${model_role}
+val_output_dir: null
+n_traces_train: 0
+n_traces_val: 5
+save_every: 20
+resume: true
+validate_final: true
+use_wandb: true
+wandb_entity: lorena-yantianyi1020
+wandb_project: monitor-cot-for-reward-hacking
+wandb_run_name: ${short_dataset_name}-${base_model}-${max_tokens}-seed_${seed}-${model_role}