JING shared-LoRA portable training bundle
Complete offline training bundle, identical to the verified S3 archive. This is not a trained LoRA checkpoint.
- Archive:
train_lora_xgen_jing.tar(184,347,228,160 bytes, about 184.35 GB / 171.69 GiB). - SHA256:
78b8ecfc4d7c21c2925d7a56da393c7e2ab9ef24c286b4d90791d9710d2861f2. - Includes JING Flash base weights, MiniMax-H3 encoding components, training code, Python 3.12 runtime, CUDA 13 toolkit and two Python environments.
- Includes 158 Evoke teacher videos, 158 official annotations, PE V10 prompts (289 case JSONs), and pre-encoded text/video caches for 149 pure-navigation cases / 587 windows.
- Shared rank-16 LoRA configuration, 20 epochs / 2980 steps. TensorBoard and offline W&B logging.
- Original model licenses and notices are preserved inside the archive. No new license is granted for third-party contents.
Restore
Install the Hugging Face CLI in your download environment, then:
hf download Wentap/train_lora_xgen_jing train_lora_xgen_jing.tar train_lora_xgen_jing.tar.sha256 --local-dir ./download
cd download
sha256sum -c train_lora_xgen_jing.tar.sha256
mkdir -p /your/path
tar -xf train_lora_xgen_jing.tar -C /your/path
cd /your/path/train_lora_xgen_jing
bash run.sh verify
bash run.sh check
bash run.sh start
bash run.sh status
The entire extracted directory can be placed under any parent path before the first run. Preserve relative symlinks. start verifies and imports all 587 encoded windows, skipping GPU encoding. The packaged OS/GPU driver is not included; use compatible Linux x86_64 and NVIDIA drivers. The current configuration expects 8 GPUs. Consult the bundled README for detailed requirements.
Validation and limitations
All 587 windows were encoded. Moving the complete directory and importing every cache succeeded; representative tensors were identical. The package's 85,132 manifest files passed SHA256 validation. The S3 archive was fully read back and matched its local SHA256.
An attempted training launch on shared 8xH100 GPUs loaded the model and caches but did not complete the first optimizer step. One rank entered exception cleanup while its GPU memory was nearly exhausted; the job was stopped. Full training, checkpoint resume and operation on the target H20 server remain unverified. This archive is the same pre-training snapshot uploaded to S3, not the later attempted run.
Teacher source: https://huggingface.co/datasets/SII-YuanyangYin/wbench_result_tubor . Prompt/video timing uses conservative windows and does not establish the original Evoke generation schedule or any WBench score improvement.