run_name: loracle_pretrain_sft_v4_sweep_A
checkpoint_dir: checkpoints/loracle_pretrain_sft_v4_sweep_A
base_model: /workspace/models/Qwen3-14B
d_model: 5120
token_dir_kind: svd_fixed_k16_mag7_rankfirst
n_direction_tokens: 4480
training_lora_dir: data/pretrain_v4/direction_tokens_svd_fixed_k16_mag7_rankfirst
prompts_path: data/pretrain_v4/prompts.parquet
loraqa_path: data/pretrain_v4/qa.parquet
holdout_ids_path: data/pretrain_v4/holdout_ids.json
interpreter_rank: 256
lora_alpha: 32
batch_size: 1
lr: 3.0e-05
warmup_steps: 500
weight_decay: 0.01
max_grad_norm: 1.0
epochs: 1
max_length: 5500
bf16: true
tasks:
- loraqa
task_weights: natural
eval_every_epochs: 0.5
cross_lora_eval_every_epochs: 0.5
early_stop_patience: 0
early_stop_min_delta: 0.0
judge_evals:
- enabled: true
  eval_set: configs/eval_sets/heldout_ia.yaml
  eval_every_epochs: 0.5
- enabled: true
  eval_set: configs/eval_sets/trigger_recovery_heldout_ia.yaml
  eval_every_epochs: 0.5
- enabled: true
  eval_set: configs/eval_sets/auditbench.yaml
  eval_every_epochs: 0.5
- enabled: true
  eval_set: configs/eval_sets/ood_models_v3.yaml
  eval_every_epochs: 0.5
judge:
  model: anthropic/claude-sonnet-4.6
  max_workers: 16
  request_timeout_s: 60
wandb_project: lora-oracles
wandb_entity: null
skip_step_0_evals: true
fineweb_lora_dir: null
fineweb_summaries_path: null
fineweb_holdout_ids_path: null
fineweb_max_train_items: null
encoder_top_k: 16
prefix_mode: rank_tagged
grad_accum_steps: 2
hook_op: add
hook_layer: 1
typed_slots: false
instruction_preamble: null
use_system_prompt: false
lr_schedule: linear