Revert "Replace use of transformers library with vLLM for high-speed inference"

Revert "Diagnosing why device not found"

Diagnosing why device not found

Replace use of transformers library with vLLM for high-speed inference

Change duration, reasoning level, max_new_tokens to ensure a smooth demo

Split out system prompt to use gpt-oss-20b's low reasoning mode, max_new_tokens->1500, comment cleanup

Update readme with awareness of ZeroGPU limitations for visitors

One last time. Duration of 45s, which should allow almost all requests to succeed. Visitors can't generate if the duration exceeds their remaining daily allotment.

Change duration to 120 (likely not a visitor setting)

Set higher ZeroGPU limit for visitors.

Create public link on launch

Address long method code smell in UI code

Populate the README, fix 'inference not run' formatting error

Some final cleanup of error messages when generation fails

Add feedback in the results section

Try to have the quiz_json box visible while the quiz is being generated.

That works functionally but the layout is bad. Have questions be inline with answer radio button groups.

Another UI rebuild. Have a fixed number of questions slots that are populated instead of trying to dynamically create UI element.

Rework how quick UI elements are populated, add A-D

Add some debugging prints for building the quiz

Additional cleanup and commenting of quiz generation code

Add empty topic check to avoid useless inference

Isolate inference from response extraction.

Extract out inference code to a separate file

Change to grabbing only text after the final instance of the start of the JSON.

Ignore 'Final Answer:' appearing during reasoning

Avoid returning CoT by finding the final answer.

Refine prompt some.

Be more explicit with instructions to remove code-generating behavior

Use roleplaying to try to influence the LLM away from writing code that generates a quiz

Fix fstring issue with json literal

Use an LLM (gpt-oss-20B) to generate quiz questions

Initial Gradio app w/ stock json

initial commit