Debug T5-Large Java Integration

The Google T5-large model compiles and runs inside my Java application, yet every call comes back with completely irrelevant answers even though the input is plain text. I’ll pass you the current Java source (Maven project) together with the prompt examples that fail.

Your job is to trace why the model misbehaves—whether it is a bad tokeniser, a wrong pipeline configuration, an inference-time parameter, or something else—and then patch the code so that identical prompts start returning sensible, context-appropriate completions. You are free to tweak hyper-parameters (top-k, temperature, max tokens), switch to the correct pre-trained checkpoint, or adjust any HuggingFace / TensorFlow Java binding I’m using, but please keep the solution inside the existing Java stack.

Deliverables
• Updated Java class / configuration files ready to compile.
• A short changelog explaining what was fixed and why.
• Proof of correct behaviour: before-and-after console logs for at least three prompts.
Model in onnx
https://huggingface.co/onnx-community/rugpt3large_based_on_gpt2-ONNX/tree/main/onnx
tokenizer model
https://huggingface.co/google-t5/t5-large/tree/main

config
https://huggingface.co/onnx-community/rugpt3large_based_on_gpt2-ONNX/blob/main/config.json

chat
https://huggingface.co/google-t5/t5-small/discussions/33

added video