2daction-stage2-checkpoint-14000

Hallucinatie/2daction-stage2-checkpoint-14000

2daction-stage2-checkpoint-14000 is a machine learning model from Hallucinatie. Use it for the machine learning task on the model card, and read the license before you ship it in a product. It is set up for pytorch.

This repository contains the public evaluation bundle for the Stage2 NPC MTP policy at training step 14,000. The checkpoint was atomically completed before the subsequent visualization-path failure, so the stored weig…

Downloads · 30 days

Access

Public

Updated Aug 25, 2026

Parameters

556M

2.6 GB on disk

Likes

Public

Hugging Face

Repo makeup

Click a slice to open those files.

.safetensors2.6 GB · 100%

Parameter types

How the weights are stored.

F32556M · 100%

Model card

From the Hugging Face model README

2DAction Stage2 checkpoint — step 14000

The matching source/data package is xixibuxixi/2daction-stage2. Extract that package according to its README before using this checkpoint.

Files

File	Purpose
`model.safetensors`	Complete `NPCMTPPolicy`: frozen video backbone, NPC branch, and structured action head.
`model_1.safetensors`	Voxel class embedding table.
`ema.safetensors`	EMA overrides for the 140 trainable NPC/action tensors. This is not a standalone model.
`load_weights.py`	Strict raw/EMA loading helper.
`test_config.yaml`	Portable architecture and evaluation settings.
`code/`	KV-cache-safe versions of the two modified Stage2 source files.
`SHA256SUMS`	SHA-256 checksums for every published payload file.

Optimizer and RNG state are intentionally omitted. This bundle supports evaluation/inference, but it is not a full training-resume checkpoint.

Setup

Download and extract the source/data package above.
Replace the corresponding files in the extracted package with the files under code/. The patch clears rollout-only KV caches in a finally block; without it, a later full-sequence forward can reuse 32-frame caches with a 65-frame training/evaluation sequence.
Instantiate MultiAgentFrameDepthStackPixelDiT and wrap it in NPCMTPPolicy(num_npc_blocks=4, horizons=8) using test_config.yaml and the original package defaults.
Load the checkpoint with the helper below.

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()

use_ema=True first loads the complete raw model and then replaces the 140 trainable tensors with their EMA values. Use False to reproduce the exact raw step-14,000 weights.

The pixel VAE, vocabulary files, policy text embeddings, validation data, and remaining model source are provided by the source/data package and are not duplicated here.

Integrity

Run this from the repository root after downloading:

sha256sum -c SHA256SUMS

Library: pytorch
Created: Aug 25, 2026
Updated: Aug 25, 2026