Spaces:
Sleeping
Sleeping
Commit History
-
Fix cache: remove cache_dir override so env.useBrowserCache=true actually works
Verified signatureba30678
-
revert to 9b50a208 + modelId only, thinking enabled
Bobbera9fb427 -
deploy: fine-tuned model with thinking disabled (Q4 thinking broken on WebGPU)
BobberTest results prove the issue is model-level: - Reference model: 124 chunks, proper </think>, content works ✓ - Our Q4 model + thinking: 1 chunk, immediate <|im_end|>, zero output ✗ - Our Q4 model + no thinking: 12 chunks, good answer ✓ The Q4 quantization breaks enable_thinking=true on WebGPU. Default to thinking disabled until we fix the ONNX export.47a7f69 -
revert: back to 9b50a208 (reference model, known working)
Bobbere446d5d -
fix: handle Q4 model missing </think> on WebGPU
BobberOn WebGPU, our Q4 model never generates the </think> token, so the think-parser stays in reasoning mode and content stays empty. When this happens, move reasoning text to content so users see the response. Changes from 9b50a208 baseline: - modelId -> bobber/lex-interviewer-nemotron-4b-grpo-v12 - eos_token_id: [2, 11] (stop at <|im_end|>) - Fallback: reasoning -> content when </think> missingab0418f -
revert: back to 9b50a208 (reference model) for testing
Bobberee3637f -
fix: MessageBubble falls back to reasoning when content is empty — eliminates 'No response'
BobberThe root issue: on WebGPU, the model sometimes doesn't output </think>, so parser.content stays empty while parser.reasoning has the full response. The LLMProvider fallback tried to fix this but due to React batching, the component could render before the state update completed. This fix is at the component level: if content is empty but reasoning exists, display reasoning as the response. This is a belt-and-suspenders fix that works regardless of state timing.e08bb06 -
debug: add FINAL console.log to trace parser state — must find root cause
Bobberd4c723d
-
fix: when model skips </think>, move reasoning to content and clear reasoning
BobberOn WebGPU the Q4 model never outputs </think>, so all text goes to reasoning and content stays empty. The fix moves reasoning to content and clears the reasoning display to avoid showing duplicate text.8802730 -
fix: extract last question from reasoning when model stops after </think> without content
BobberOn WebGPU, Q4 precision differences can cause the model to output </think> then immediately <|im_end|> with no answer text. The fallback extracts the last question from the reasoning block (since the model was prompted to ask questions).4de85d0 -
fix: remove cache_dir override — use default browser Cache API for ONNX persistence
Bobberdc12d71 -
fix: handle missing </think> — fallback reasoning to content, never show 'No response'
Bobber9b9b289 -
debug: add console logging to diagnose 'No response' issue
Bobber4eb33dc -
fix: explicit eos_token_id=[2,11] at generation time — works even with cached model config
Bobber9446127 -
fix: revert to reference streamer approach + fix eos_token_id in model config
Bobber8ae6647 -
fix: filter special tokens and trailing whitespace from stream output
Bobber6572f10 -
fix: pin ort-web to 1.16.3 to fix Module.MountedFiles error with q4 ONNX external data
BobberThe @huggingface/transformers 4.0.0-next.8 pulls ort-web 1.22.x-dev which removed the Module.MountedFiles API used for loading external .onnx_data files. This caused: Error: Module.MountedFiles is not available Fix: downgrade to @huggingface/transformers 3.5.0 and override onnxruntime-web to 1.16.3 (last version with working MountedFiles for external data loading). Also removes bundled WASM (now served from jsDelivr CDN at runtime) and uses stable vite 5.x instead of beta 8.x.3f50c26
-
Enable browser Cache API for ONNX files, better download progress
Verified signaturefc04968 -
Add settings panel with editable system prompt, update examples
Verified signature41b224e
-
Customize for Lex Fridman interviewer: system prompt, examples, branding
Verified signature4e22e37
-
Upload 88 files
Verified signaturef672a5d
