# A12 — Init logit = 3.0 (moderate, sigmoid not saturated)
# Purpose: A starts at σ(1.5)≈0.82, gradient ~250× larger than logit=15.
#          Does predictor learn useful topology when sigmoid is not saturated?
# Run: python scripts/train.py --config configs/a12_init_logit_3.yaml

# Model
olmo_model_id: "allenai/OLMo-2-0425-1B"
qwen_model_id: "Qwen/Qwen3-Embedding-0.6B"

# Predictor
predictor_hidden_dim: 1024
predictor_rank: 32
cascading_gate_k: 5.0
input_norm: "none"
init_logit: 3.0

# Data
dataset: "allenai/dolma"
dataset_name: "v1_7"
seq_len: 1024
batch_size: 4
micro_batch_size: 2
qwen_input_prefix: ""

# Eval
eval_skip: 10000
eval_size: 50

# Training — ~50M tokens = 12500 steps
total_steps: 12500
lr: 3e-4
weight_decay: 0.01
optimizer: "adamw"

# Schedules — constant tau=2 (same as S2), no sparsity
tau_init: 2.0
tau_final: 2.0
tau_schedule: "cosine"
lambda_max: 0.0
lambda_warmup_frac: 0.2

# Logging
wandb_project: "dagformer"
wandb_run_name: "a12-init-logit-3"
log_every: 10
eval_every: 500

# Checkpointing
save_every: 2500
save_dir: "checkpoints/a12/"

# Hardware
num_gpus: 1