Release 4D-R1 GRPO and NFT LoRA adapters with mixed-precision variants

4D-R1-GRPO-MP/README.md

@@ -0,0 +1,9 @@
+# 4D-R1-GRPO-MP
+
+4D-R1 LoRA adapter corresponding to **GRPO (mixed-precision)** in the paper.
+
+Use this adapter with the VerseCrafter base model and the 4D-R1 inference implementation. This directory contains adapter weights and configuration, not a standalone base model.
+
+LoRA rank: 16. Target modules: q, k, v, o.
+
+MP denotes mixed-precision training; it does not mean that the saved adapter is an FP4-quantized standalone model.

4D-R1-GRPO-MP/adapter_config.json

@@ -0,0 +1,50 @@
+{
+  "alora_invocation_tokens": null,
+  "alpha_pattern": {},
+  "arrow_config": null,
+  "auto_mapping": {
+    "base_model_class": "VerseCrafterWanTransformer3DModel",
+    "parent_library": "versecrafter.models.wan_transformer3d_versecrafter"
+  },
+  "base_model_name_or_path": null,
+  "bias": "none",
+  "corda_config": null,
+  "ensure_weight_tying": false,
+  "eva_config": null,
+  "exclude_modules": null,
+  "fan_in_fan_out": false,
+  "inference_mode": true,
+  "init_lora_weights": true,
+  "layer_replication": null,
+  "layers_pattern": null,
+  "layers_to_transform": null,
+  "loftq_config": {},
+  "lora_alpha": 16,
+  "lora_bias": false,
+  "lora_dropout": 0.0,
+  "lora_ga_config": null,
+  "megatron_config": null,
+  "megatron_core": "megatron.core",
+  "modules_to_save": null,
+  "monteclora_config": null,
+  "peft_type": "LORA",
+  "peft_version": "0.20.0",
+  "qalora_group_size": 16,
+  "r": 16,
+  "rank_pattern": {},
+  "revision": null,
+  "target_modules": [
+    "v",
+    "k",
+    "q",
+    "o"
+  ],
+  "target_parameters": null,
+  "task_type": null,
+  "trainable_token_indices": null,
+  "use_bdlora": null,
+  "use_dora": false,
+  "use_qalora": false,
+  "use_rslora": false,
+  "velora_config": null
+}

4D-R1-GRPO-MP/adapter_model.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:d2d67e6260d24f2126d2c258e3e2818c731034719eb4e2920c278b3feb5298d0
+size 251755360

4D-R1-GRPO/README.md

@@ -0,0 +1,7 @@
+# 4D-R1-GRPO
+
+4D-R1 LoRA adapter corresponding to **GRPO** in the paper.
+
+Use this adapter with the VerseCrafter base model and the 4D-R1 inference implementation. This directory contains adapter weights and configuration, not a standalone base model.
+
+LoRA rank: 16. Target modules: q, k, v, o.

4D-R1-GRPO/adapter_config.json

@@ -0,0 +1,48 @@
+{
+  "alora_invocation_tokens": null,
+  "alpha_pattern": {},
+  "arrow_config": null,
+  "auto_mapping": {
+    "base_model_class": "VerseCrafterWanTransformer3DModel",
+    "parent_library": "versecrafter.models.wan_transformer3d_versecrafter"
+  },
+  "base_model_name_or_path": null,
+  "bias": "none",
+  "corda_config": null,
+  "ensure_weight_tying": false,
+  "eva_config": null,
+  "exclude_modules": null,
+  "fan_in_fan_out": false,
+  "inference_mode": true,
+  "init_lora_weights": true,
+  "layer_replication": null,
+  "layers_pattern": null,
+  "layers_to_transform": null,
+  "loftq_config": {},
+  "lora_alpha": 16,
+  "lora_bias": false,
+  "lora_dropout": 0.0,
+  "lora_ga_config": null,
+  "megatron_config": null,
+  "megatron_core": "megatron.core",
+  "modules_to_save": null,
+  "peft_type": "LORA",
+  "peft_version": "0.19.1",
+  "qalora_group_size": 16,
+  "r": 16,
+  "rank_pattern": {},
+  "revision": null,
+  "target_modules": [
+    "v",
+    "k",
+    "q",
+    "o"
+  ],
+  "target_parameters": null,
+  "task_type": null,
+  "trainable_token_indices": null,
+  "use_bdlora": null,
+  "use_dora": false,
+  "use_qalora": false,
+  "use_rslora": false
+}

4D-R1-GRPO/adapter_model.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:f3e4133c0586ff1241ca143641086913964fe1bc76ee17ba3f153544d786c4ec
+size 251755360

4D-R1-NFT-MP/README.md

@@ -0,0 +1,9 @@
+# 4D-R1-NFT-MP
+
+4D-R1 LoRA adapter corresponding to **NFT (mixed-precision)** in the paper.
+
+Use this adapter with the VerseCrafter base model and the 4D-R1 inference implementation. This directory contains adapter weights and configuration, not a standalone base model.
+
+LoRA rank: 16. Target modules: q, k, v, o.
+
+MP denotes mixed-precision training; it does not mean that the saved adapter is an FP4-quantized standalone model.

4D-R1-NFT-MP/adapter_config.json

@@ -0,0 +1,50 @@
+{
+  "alora_invocation_tokens": null,
+  "alpha_pattern": {},
+  "arrow_config": null,
+  "auto_mapping": {
+    "base_model_class": "VerseCrafterWanTransformer3DModel",
+    "parent_library": "versecrafter.models.wan_transformer3d_versecrafter"
+  },
+  "base_model_name_or_path": null,
+  "bias": "none",
+  "corda_config": null,
+  "ensure_weight_tying": false,
+  "eva_config": null,
+  "exclude_modules": null,
+  "fan_in_fan_out": false,
+  "inference_mode": true,
+  "init_lora_weights": true,
+  "layer_replication": null,
+  "layers_pattern": null,
+  "layers_to_transform": null,
+  "loftq_config": {},
+  "lora_alpha": 16,
+  "lora_bias": false,
+  "lora_dropout": 0.0,
+  "lora_ga_config": null,
+  "megatron_config": null,
+  "megatron_core": "megatron.core",
+  "modules_to_save": null,
+  "monteclora_config": null,
+  "peft_type": "LORA",
+  "peft_version": "0.20.0",
+  "qalora_group_size": 16,
+  "r": 16,
+  "rank_pattern": {},
+  "revision": null,
+  "target_modules": [
+    "k",
+    "q",
+    "o",
+    "v"
+  ],
+  "target_parameters": null,
+  "task_type": null,
+  "trainable_token_indices": null,
+  "use_bdlora": null,
+  "use_dora": false,
+  "use_qalora": false,
+  "use_rslora": false,
+  "velora_config": null
+}

4D-R1-NFT-MP/adapter_model.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:8163548a7b8e268ed52bfa2144c19bfe45692b436a7c3d37908db306b1119b89
+size 251755360

4D-R1-NFT/README.md

@@ -0,0 +1,7 @@
+# 4D-R1-NFT
+
+4D-R1 LoRA adapter corresponding to **NFT** in the paper.
+
+Use this adapter with the VerseCrafter base model and the 4D-R1 inference implementation. This directory contains adapter weights and configuration, not a standalone base model.
+
+LoRA rank: 16. Target modules: q, k, v, o.

4D-R1-NFT/adapter_config.json

@@ -0,0 +1,48 @@
+{
+  "alora_invocation_tokens": null,
+  "alpha_pattern": {},
+  "arrow_config": null,
+  "auto_mapping": {
+    "base_model_class": "VerseCrafterWanTransformer3DModel",
+    "parent_library": "versecrafter.models.wan_transformer3d_versecrafter"
+  },
+  "base_model_name_or_path": null,
+  "bias": "none",
+  "corda_config": null,
+  "ensure_weight_tying": false,
+  "eva_config": null,
+  "exclude_modules": null,
+  "fan_in_fan_out": false,
+  "inference_mode": true,
+  "init_lora_weights": true,
+  "layer_replication": null,
+  "layers_pattern": null,
+  "layers_to_transform": null,
+  "loftq_config": {},
+  "lora_alpha": 16,
+  "lora_bias": false,
+  "lora_dropout": 0.0,
+  "lora_ga_config": null,
+  "megatron_config": null,
+  "megatron_core": "megatron.core",
+  "modules_to_save": null,
+  "peft_type": "LORA",
+  "peft_version": "0.19.1",
+  "qalora_group_size": 16,
+  "r": 16,
+  "rank_pattern": {},
+  "revision": null,
+  "target_modules": [
+    "q",
+    "o",
+    "v",
+    "k"
+  ],
+  "target_parameters": null,
+  "task_type": null,
+  "trainable_token_indices": null,
+  "use_bdlora": null,
+  "use_dora": false,
+  "use_qalora": false,
+  "use_rslora": false
+}

4D-R1-NFT/adapter_model.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:fa0bc93cacf20b38cd083e6bda7504f3ca3ad23abf6670fa2be0e5e80a04d3cc
+size 251755360

README.md

@@ -0,0 +1,28 @@
+---
+library_name: peft
+tags:
+- lora
+- video-generation
+- motion-control
+---
+
+# 4D-R1
+
+LoRA adapters for 4D-R1, a reinforcement post-training framework for motion control and efficient video generation built on VerseCrafter.
+
+| Directory | Paper model |
+|---|---|
+| `4D-R1-GRPO` | GRPO |
+| `4D-R1-NFT` | NFT |
+| `4D-R1-GRPO-MP` | GRPO (mixed-precision) |
+| `4D-R1-NFT-MP` | NFT (mixed-precision) |
+
+Each directory contains `adapter_model.safetensors` and `adapter_config.json`. The adapters use LoRA rank 16 on the q, k, v, and o projections.
+
+These are LoRA adapters, not standalone full-model checkpoints. Inference requires the VerseCrafter base model, its companion components, and the compatible 4D-R1 inference implementation. Base-model weights are not included in this repository.
+
+MP denotes mixed-precision training. It does not indicate that the saved adapter weights form an FP4-quantized model.
+
+The corresponding paper rows are GRPO, NFT, GRPO (mixed-precision), and NFT (mixed-precision). Internal run identifiers and update counts are not part of the public checkpoint names.
+
+`SHA256SUMS` provides checksums for the adapter weights and configuration files.

SHA256SUMS

@@ -0,0 +1,8 @@
+f3e4133c0586ff1241ca143641086913964fe1bc76ee17ba3f153544d786c4ec  4D-R1-GRPO/adapter_model.safetensors
+e96450a12b06910cab904f0af4341ceb95ae458fdc0b574fbed91564efecabfd  4D-R1-GRPO/adapter_config.json
+fa0bc93cacf20b38cd083e6bda7504f3ca3ad23abf6670fa2be0e5e80a04d3cc  4D-R1-NFT/adapter_model.safetensors
+86cc78cdf02076d5f9528aed52da0fd4b45ab99578ca4c0a6a6599229c5d78d9  4D-R1-NFT/adapter_config.json
+d2d67e6260d24f2126d2c258e3e2818c731034719eb4e2920c278b3feb5298d0  4D-R1-GRPO-MP/adapter_model.safetensors
+7170c477b8f0ad76fad4d93f6f6d45935fe362bfca89eaf39943d06ee76c9308  4D-R1-GRPO-MP/adapter_config.json
+8163548a7b8e268ed52bfa2144c19bfe45692b436a7c3d37908db306b1119b89  4D-R1-NFT-MP/adapter_model.safetensors
+234edf984526b60e42ab13766fadf94eefdba2a18928619c4e991c3051dc35fa  4D-R1-NFT-MP/adapter_config.json