# ==========================================================================
# Training args for the released anchor SFT run.
# PHASE-Tree hyper-LoRA SFT on Qwen2.5-7B-Instruct.
# Variant: warm-start, lr=5e-6 (anchor SFT run).
# Effective batch size = batch_size * grad_accum_steps = 4 * 2 = 8.
# 40000 optimizer steps; a checkpoint is saved every val_freq=5000 steps.
# ==========================================================================
run_name: anchor
save_dir: phase_tree_models/sft/hyper_lora

# ── Model ────────────────────────────────────────────────────────────────
model_dir: Qwen/Qwen2.5-7B-Instruct
emb_model: Qwen/Qwen3-Embedding-4B
init_hypermod_from: phase_tree_models/phase_tree_pretrained/hypermod.pt
init_hypermod_strict: false

# ── Task / data ──────────────────────────────────────────────────────────
training_task: sft
exp_setup: hyper_lora
sft_mode: completion
encoder_type: linear
train_ds_names:
- RAIDEN_train
- CharacterEval_train
- HPD_train
- SimsConv_train
- ChatHaruhi_train
- Friends_train
- StarTrek_TNG_train
- TheOffice_train
eval_ds_info:
- RAIDEN_random_test
- RAIDEN_ood_test
- CharacterEval_random_test
- CharacterEval_ood_test
- HPD_random_test
- HPD_ood_test
- SimsConv_random_test
- SimsConv_ood_test
- ChatHaruhi_random_test
- ChatHaruhi_ood_test
- Friends_random_test
- Friends_ood_test
- StarTrek_TNG_random_test
- StarTrek_TNG_ood_test
- TheOffice_random_test
- TheOffice_ood_test
use_per_task_emb: true
use_one_hot_task_emb: false
use_inp_as_desc: false
use_per_sample_desc: false
use_default_desc: false
use_hierarchical_sampler: true
dataset_sampling_strategy: sqrt_size
equally_weight_sample: true
n_tasks_per_batch: 6
n_points_per_task: 2
inp_max_len: 1024
target_modules:
- q_proj
- v_proj

# ── Hypermod architecture ────────────────────────────────────────────────
use_hypernet: true
head_in_size: 2048
head_use_bias: false
hypernet_latent_size: 1024
delta_w_scaling: 100
pred_z_score: true
factorized: false
shared_AB_head: false
autoreg_gen: false
learnable_pos_emb: false
learnable_AB_offset: false
freeze_heads: false

# ── Fusion (disabled for this run, kept for loader compatibility) ────────
use_conv_fusion: false
conv_fusion_type: 1d
conv_fusion_kernel_size: 3
conv_fusion_num_layers: 2
conv_fusion_channels: 64
conv_fusion_dropout: 0.1
use_attention_fusion: false
attention_fusion_type: self
attention_num_heads: 8
attention_num_layers: 2
attention_dropout: 0.1

# ── Optimisation ─────────────────────────────────────────────────────────
lr: 5.0e-06
weight_decay: 0.01
warmup_frac: 0.05
max_grad_norm: 1.0
label_smoothing: 0.1
l2_reg_generated_w: 0.001
neftune_noise_alpha: 5.0
gradient_checkpointing: true

# ── Schedule / batching ──────────────────────────────────────────────────
epochs: 40000
batch_size: 4
grad_accum_steps: 2
val_batch_size: 16
seed: 42

# ── Logging / saving ─────────────────────────────────────────────────────
logging_freq: 50
val_freq: 5000                # validation + checkpoint cadence (steps)
top_k_checkpoints: 999        # keep every checkpoint
skip_val: true                # validation done post-hoc via eval scripts
skip_eval: false
use_early_stopping: true
early_stopping_patience: 5
early_stopping_min_delta: 0.0