# S2 — Gradient flow check (constant tau=2, ~50M tokens)
# Purpose: Lower tau gives sharper gates. Does predictor learn useful topology?
#          Any NLL drop below baseline (2.4569) = gradient flows correctly.
# Run: python scripts/train.py --config configs/s2_gradient_flow.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"

# 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 @ batch=4, seq=1024
total_steps: 12500
lr: 3e-4
weight_decay: 0.01
optimizer: "adamw"

# Schedules — constant tau=2 (sharper gates than S1), 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: "s2-gradient-flow"
log_every: 10
eval_every: 500

# Checkpointing
save_every: 2500
save_dir: "checkpoints/s2/"

# Hardware
num_gpus: 1