| # ========================================================================== | |
| # Architecture / loading metadata for the PHASE-Tree pretrained hypermod. | |
| # | |
| # The released warm-start checkpoint is | |
| # phase_tree_models/phase_tree_pretrained/hypermod.pt | |
| # (= the it_20000 snapshot of the original pretraining run). | |
| # | |
| # Only fields read by `load_hypermod_checkpoint` (path resolution + | |
| # hypermod architecture) are kept here; the dataset lists and the | |
| # original training schedule are intentionally omitted, because the | |
| # PHASE-Tree SFT runs warm-start from these weights and override every | |
| # training hyperparameter from `train_phase_tree_qwen_7b.sh`. | |
| # ========================================================================== | |
| # ── Paths ──────────────────────────────────────────────────────────────── | |
| model_dir: Qwen/Qwen2.5-7B-Instruct | |
| emb_model: Qwen/Qwen3-Embedding-4B | |
| mt_lora_path: null | |
| # ── Task setup ─────────────────────────────────────────────────────────── | |
| training_task: sft | |
| exp_setup: hyper_lora | |
| sft_mode: completion | |
| encoder_type: linear | |
| # ── Task-embedding mode ───────────────────────────────────────────────── | |
| use_hypernet: true | |
| 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 | |
| # ── Hypermod architecture ──────────────────────────────────────────────── | |
| 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 | |
| # ── Fusion (disabled; 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 | |
| # ── Target LoRA modules and context window ────────────────────────────── | |
| target_modules: | |
| - q_proj | |
| - v_proj | |
| inp_max_len: 1024 | |