Add PetGPT-SMPL model weights

model_weights/petgpt_smpl/CHECKPOINT_METADATA.json

@@ -0,0 +1,38 @@
+{
+  "pet_gpt_smpl_best.pt": {
+    "best_eval_loss": 3.106193939844767,
+    "sha256": "3d946e3bdb156d4a106e02a277ea7b4bbc84ccf291f6ec6189ffbe0bc3f589ca",
+    "size_bytes": 121658083,
+    "source_role": "pet_gpt_smpl"
+  },
+  "pet_motion_classifier_best.pt": {
+    "best_eval_acc": 0.9642857142857143,
+    "sha256": "9b015fe7f3074e663a4d1f75f100abaf25d3f0efb696256f2ccfff8cd0208fe7",
+    "size_bytes": 41467349,
+    "source_role": "pet_motion_classifier"
+  },
+  "wan_mano_vqvae_best.pt": {
+    "best_eval_loss": 1.7700489613474633,
+    "sha256": "36d8933b5d51862affcbe4b29851298aa6b144a6d46da9436ba911a66fd227bc",
+    "size_bytes": 349270763,
+    "source_role": "wan_mano_vqvae"
+  },
+  "wan_pet_rel_vqvae_best.pt": {
+    "best_eval_loss": 0.11038906753428875,
+    "sha256": "e398b191436d500c9991ddc003424b9875a9441be424abe90b45f19c8bc93681",
+    "size_bytes": 108912982,
+    "source_role": "wan_pet_rel_vqvae"
+  },
+  "wan_pet_vqvae_global_best.pt": {
+    "best_eval_loss": 0.27014976891420656,
+    "sha256": "9c4f9952acf570b4491b5107b7f8f6fbd0b2d20fc2bd2d4953fabf5f0c5d6d3c",
+    "size_bytes": 109135051,
+    "source_role": "wan_pet_vqvae_global"
+  },
+  "wan_smpl_vqvae_best.pt": {
+    "best_eval_loss": 0.8021610458286441,
+    "sha256": "89e28e7e36a75f7a2492833cc1807163f76d921185ad8ebb97a69f102844eb81",
+    "size_bytes": 174627353,
+    "source_role": "wan_smpl_vqvae"
+  }
+}

model_weights/petgpt_smpl/MODEL_CARD.md

@@ -0,0 +1,57 @@
+# Model card: InterPet4D PetGPT-SMPL
+
+## Model description
+
+PetGPT-SMPL is an autoregressive prefix language model over frozen motion VQ
+tokens. It uses left/right MANO motion, SMPL body motion, and aligned MERT
+audio features as conditioning and predicts pet-relative-position tokens and
+pet-motion tokens.
+
+## Intended use
+
+The model is intended for research on human-pet motion generation, benchmark
+reproduction, ablation studies, and qualitative visualization. It is not
+validated for safety-critical, veterinary, biometric, surveillance, or
+real-time control applications.
+
+## Training and evaluation data
+
+The matched PetGPT-SMPL subset contains 182 sequences split deterministically
+by dog identity into 146 training and 36 validation files with seed 42. The
+release includes filenames only; it does not redistribute the underlying pet,
+MANO, SMPL, audio, or participant data.
+
+WanPetRelVQVAE, WanManoVQVAE, and WanSmplVQVAE use the same 146/36 aligned
+split. WanPetVQVAE is trained on the full 227-sequence pet-motion set, split
+into 187 training and 40 validation files. Exact train/test filenames for both
+split families are included under `splits/`.
+
+## Evaluation settings
+
+- random seed: 42
+- sampling temperature: 1.0
+- top-k: 50
+- classifier checkpoint accuracy: 0.9642857
+- PetGPT FID: 5.00918468
+- VQ reconstruction FID: 0.62121353
+
+Compare results only when classifier weights, preprocessing, and split match.
+
+## Limitations
+
+- Evaluation is based on a small, identity-structured dataset.
+- Autoregressive sampling is stochastic even with fixed hyperparameters; the
+  released evaluator fixes all available random seeds for reproducibility.
+- The model may generate implausible poses, temporal jitter, identity drift,
+  or motion that is weakly aligned with the human conditioning signal.
+- Performance outside the released dog identities, capture setup, skeleton
+  definition, frame rates, or normalization statistics is unknown.
+- The model has not been evaluated for demographic, participant, or animal
+  representation bias.
+
+## Privacy and redistribution
+
+No raw participant data, audio, motion sequences, W&B logs, usernames, API
+keys, or internal filesystem paths are included in the release package.
+Confirm that the dataset consent and redistribution terms permit the intended
+public use before publishing any example inputs or outputs.

model_weights/petgpt_smpl/README.md

@@ -0,0 +1,19 @@
+# PetGPT-SMPL weights
+
+This directory contains the six weights-only checkpoints for PetGPT-SMPL. The
+corresponding evaluation code is available at:
+
+https://github.com/puckikk1202/interpet4d_post
+
+The released evaluator reports a PetGPT FID of 5.0092 and a VQ reconstruction
+FID of 0.6212.
+
+Run the following from the code repository to download and verify every file:
+
+```bash
+python download_checkpoints.py
+python smoke_test.py
+```
+
+Individual SHA-256 hashes and validation metadata are recorded in
+`CHECKPOINT_METADATA.json` and `SHA256SUMS`.

model_weights/petgpt_smpl/SHA256SUMS

@@ -0,0 +1,6 @@
+3d946e3bdb156d4a106e02a277ea7b4bbc84ccf291f6ec6189ffbe0bc3f589ca  checkpoints/pet_gpt_smpl_best.pt
+9b015fe7f3074e663a4d1f75f100abaf25d3f0efb696256f2ccfff8cd0208fe7  checkpoints/pet_motion_classifier_best.pt
+36d8933b5d51862affcbe4b29851298aa6b144a6d46da9436ba911a66fd227bc  checkpoints/wan_mano_vqvae_best.pt
+e398b191436d500c9991ddc003424b9875a9441be424abe90b45f19c8bc93681  checkpoints/wan_pet_rel_vqvae_best.pt
+9c4f9952acf570b4491b5107b7f8f6fbd0b2d20fc2bd2d4953fabf5f0c5d6d3c  checkpoints/wan_pet_vqvae_global_best.pt
+89e28e7e36a75f7a2492833cc1807163f76d921185ad8ebb97a69f102844eb81  checkpoints/wan_smpl_vqvae_best.pt

model_weights/petgpt_smpl/checkpoints/pet_gpt_smpl_best.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:3d946e3bdb156d4a106e02a277ea7b4bbc84ccf291f6ec6189ffbe0bc3f589ca
+size 121658083

model_weights/petgpt_smpl/checkpoints/pet_motion_classifier_best.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:9b015fe7f3074e663a4d1f75f100abaf25d3f0efb696256f2ccfff8cd0208fe7
+size 41467349

model_weights/petgpt_smpl/checkpoints/wan_mano_vqvae_best.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:36d8933b5d51862affcbe4b29851298aa6b144a6d46da9436ba911a66fd227bc
+size 349270763

model_weights/petgpt_smpl/checkpoints/wan_pet_rel_vqvae_best.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:e398b191436d500c9991ddc003424b9875a9441be424abe90b45f19c8bc93681
+size 108912982

model_weights/petgpt_smpl/checkpoints/wan_pet_vqvae_global_best.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:9c4f9952acf570b4491b5107b7f8f6fbd0b2d20fc2bd2d4953fabf5f0c5d6d3c
+size 109135051

model_weights/petgpt_smpl/checkpoints/wan_smpl_vqvae_best.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:89e28e7e36a75f7a2492833cc1807163f76d921185ad8ebb97a69f102844eb81
+size 174627353