Hallucinatie/2daction-stage2-checkpoint-60000

2DAction Stage2 checkpoint 60000 (grad step 30000)

Public evaluation bundle for the Stage2 NPC MTP policy at trainer step 60,000, which equals optimizer/gradient step 30,000 because this continuation used grad_accumulation_steps=2.

The checkpoint belongs to the continuation run stage2_npc_text_mtp-retrain-from-grad14000-20260825T022321Z. It was loaded back successfully with 454 model tensors, 140 optimizer states, step 60,000, grad step 30,000, and learning rate 1e-4. Only evaluation files are published here; optimizer and RNG state are intentionally omitted.

The matching source/data package is xixibuxixi/2daction-stage2.

Contents

• model.safetensors: complete NPCMTPPolicy (video backbone, NPC branch, action head).
• model_1.safetensors: voxel class embedding table.
• ema.safetensors: EMA overrides for the 140 trainable NPC/action tensors.
• load_weights.py: strict raw/EMA loading helper.
• test_config.yaml: portable architecture and evaluation settings.
• code/: KV-cache-safe Stage2 source files.
• SHA256SUMS: checksums for every payload file.

Evaluation

Extract the source/data package, replace its corresponding files with code/, and instantiate NPCMTPPolicy(num_npc_blocks=4, horizons=8) using the original package defaults plus test_config.yaml.

from load_weights import load_stage2_checkpoint

load_stage2_checkpoint(
    policy,
    voxel_class_embedder,
    checkpoint_dir="/path/to/this/repository",
    use_ema=True,
)
policy.eval()

EMA is the recommended evaluation mode. The helper first loads the complete raw model and then overlays EMA tensors. Set use_ema=False for the exact raw step-60,000 weights. The pixel VAE, vocabulary files, policy text embeddings, validation data, and remaining source files come from the source/data package.

Verify a download with:

sha256sum -c SHA256SUMS