add mult ablation arcu_1gev

checkpoints/ablations/mult/arcu_1gev/mult_default.yaml

@@ -0,0 +1,93 @@
+# Multiplicity model training config.
+# Usage: python scripts/train.py configs/mult_default.yaml
+#
+# Amortized Ar+Cu multiplicity model at a 1 GeV gun: clone of neurips_2026/arcu,
+# retrained on rp_arcu_1gev_240726. Conditions on Z/A already, so only run.name /
+# run.ckpt_dir and dl_conf.lds_args.data change -- the mult model has no
+# max_energy of its own (it reads the normalised tokens the dataset was prepped
+# with), so nothing else follows from the higher gun energy.
+
+run:
+  train_model: mult          # fm | mult
+  name: rp_mult_arcu_1gev_240726
+  devices: [0, 1, 2, 3]
+  precision: 32
+  matmul_precision: medium
+  dtype: float32
+  epochs: 10
+  dataset_size: 20_000_000     # only used to compute scheduler total_steps (2 materials x 500 items x 20k)
+  compile: false             # only applies to the fm model
+  gradient_clip_val: 1.0
+  val_check_interval: 0.2
+  limit_val_batches: 4
+  strategy: ddp
+  ckpt_dir: /ptmp/mpp/hildebra/lego/models/mult/neurips_2026/arcu_1gev   # explicit dir -> <name>.pt saved directly here (null -> $LEGO_CKPT_DIR -> ./checkpoints/ + flow|mult)
+
+logging:
+  comet: true
+  project: lego_pdgid        # api key / workspace come from .env
+
+config:
+  dl_conf:
+    lds_args:
+      data: neurips_2026/rp_arcu_1gev_240726   # prefixed with $LEGO_DATA_DIR at runtime
+      frac: 1.0
+      cutoff_mev: 10
+      min_particles: 0
+    bs: 4096
+    num_workers: 16
+
+  mm_conf:
+    in_dim: 11
+    cond_scalars: ["Density", "Z", "A", "Size"]
+    dropout: 0.05
+    h_dim: 128
+    n_layers: 6
+    n_heads: 6
+    bs: 4096
+    lr: 1.0e-3
+    weight_decay: 1.0e-2
+    warmup_steps: 0
+    post_emb_norm: false
+    use_abs_pos_emb: true
+    train_inverse: false       # co-train the separate inverse-PID model (InvModel)
+    model_args:                # splatted into the count model's Decoder
+      ff_swish: true
+      ff_glu: true
+      attn_qk_norm: true
+      rotary_pos_emb: true
+      use_adaptive_rmsnorm: true
+      use_adaptive_layerscale: true
+      residual_attn: true
+    # inverse model (used only when train_inverse: true); inv_h_dim/_n_layers/
+    # _n_heads default to the count values, inv_model_args defaults to model_args.
+    # Note: rotary_xpos is invalid on the bidirectional Encoder.
+    inv_model_args:            # splatted into the inverse model's Encoder
+      ff_swish: true
+      ff_glu: true
+      attn_qk_norm: true
+      rotary_pos_emb: true
+      use_adaptive_rmsnorm: true
+      use_adaptive_layerscale: true
+      residual_attn: true
+
+  opt_conf:
+    opt: muon
+    lr: 1.0e-2
+    momentum: 0.95
+    nesterov: true
+    ns_steps: 5
+    weight_decay: 7.0e-3
+    weight_decouple: true
+    adamw_lr: 3.0e-3
+    adamw_betas: [0.9, 0.999]
+    adamw_wd: 1.0e-2
+    adamw_eps: 1.0e-8
+    scheduler:                # total_steps computed from run.* when absent
+      cls: warmup_cosine
+      warmup_frac: 0.1
+      eta_min: 1.0e-6
+      interval: step
+
+  additional:
+    notes: "Amortized over argon + copper at a 1 GeV gun (rp_arcu_1gev_240726)."

checkpoints/ablations/mult/arcu_1gev/rp_mult_arcu_1gev_240726.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:edf54be452c543dc9dc3ffb849973e5cc11fdc9c7e01d1ba9a7ce55e7765309e
+size 11435913