Switch to greedy decoding for deterministic, clean LQ outputsβ
Nhughes09
96285e9
- Remove temperature/top_k sampling (multinomial)
- Use argmax for deterministic token selection
- Aligns with local probe methodology
- Reduces jitter and garbled character noise
Deploy V8 Reasoning Agent β 8/8 comprehension, zero 11.72 leaksβ
Nhughes09
9777e17
- Update app.py to load v8_reasoning_agent.pt by default
- Copy fixed char_tokenizer.json (space at ID 100, 245 tokens)
- Add V8 model (44.6 MB, EliteWitnessV5 fine-tuned on 19K reasoning+agent pairs)
- Ready for Hugging Face Space deployment
Phase 2: Fix tokenizer decode, upgrade reasoning trace with glyph-by-glyph analysis, add English summary, fix Gradio compatibilityβ
Nhughes09
8bea29b
- Decode now skips UNK tokens (ID 3) to prevent <unk> in output
- Added build_reasoning_trace(): parses LQ glyphs and looks up meanings from reasoning_glyphs.json
- Added generate_summary(): produces natural English description of LQ program
- Updated translate() to use new trace + summary, formatted output
- Kept ChatInterface with (message, history) signature
Fix translate() signature for Gradio ChatInterface compatibility (add history parameter)
Nhughes09
fb37a54
Remove incompatible theme parameter from ChatInterface (Gradio compatibility fix)
Nhughes09
5c73b5a
Fix syntax error in default reasoning f-string (line 169)
Nhughes09
9a799ab
Add Hugging Face Space metadata block (title, emoji, colors, sdk, app_file, pinned)
