loracle_config.yaml

run_name: kl_distill_N01000_v2
checkpoint_dir: /workspace/persona_v4/kl_distill/N01000_v2/checkpoints
base_model: Qwen/Qwen3-14B
d_model: 5120
token_dir_kind: svd_fixed_k16_mag7_rankfirst
n_direction_tokens: 4480
prefix_mode: rank_tagged
training_lora_dir: /workspace/persona_v4/kl_distill/N01000_v2/tokens
prompts_path: /workspace/persona_v4/kl_distill/N01000_v2/prompts.parquet
loraqa_path: /workspace/persona_v4/kl_distill/N01000_v2/qa.parquet
holdout_ids_path: /workspace/persona_v4/kl_distill/N01000_v2/holdout.json
interpreter_rank: 256
lora_alpha: 32
batch_size: 1
lr: 1.0e-05
warmup_steps: 200
weight_decay: 0.01
max_grad_norm: 1.0
epochs: 1
max_length: 5500
bf16: true
tasks:
- loraqa
task_weights: natural
eval_every_epochs: 0.1
cross_lora_eval_every_epochs: 0.1
early_stop_patience: 0
early_stop_min_delta: 0.0
judge_evals: []
judge:
  model: anthropic/claude-sonnet-4.6
  max_workers: 16
  request_timeout_s: 60
wandb_project: lora-oracles
wandb_entity: null
quant_4bit: false
optim_8bit: false
fineweb_lora_dir: null
fineweb_summaries_path: null
fineweb_holdout_ids_path: null
fineweb_max_train_items: null
init_interpreter_adapter: /workspace/lol_data/v7_ckpt/interpreter
save_full_trainer_state: false
init_full_trainer_state: null
seed: 42
skip_intermittent_evals: false
encoder_type: ao
encoder_top_k: 16
per_k_affines: false
learned_scale_layout: all14_rankfirst
learned_scale_init: 2.0
learned_scale_lr_mult: 50.0
learned_scale_hook_mode: modulated
hook_op: add
hook_layer: 1
hook_mode: norm_match
grad_accum_steps: 8
skip_step_0_evals: true
typed_slots: false
instruction_preamble: null
use_system_prompt: false
lr_schedule: linear
attn_implementation: sdpa
checkpoint_every: 1
ddp_find_unused_parameters: false
ddp_static_graph: true
ddp_gradient_as_bucket_view: true
sft_checkpoint: null
encoder_ckpt: null
pairs_parquet: null
kto_pairs_path: null
kto_beta: 0.1
kto_desirable_weight: 1.0
kto_undesirable_weight: 1.0
dpo_beta: 0.1