Spaces:
Running
Running
Commit History
4 days ago
-
fix Running
b91f21d
29 days ago
-
Adopt Panda-inspired comparison interface
b828a40
-
better visualization
7a459aa
-
improvements
48754a8 -
bug fix and improvements
24da7ec -
Checkpoint before tooltip palette update
350392a
-
bug fix
25a9607 -
bug fix and improvements
8f8c689 -
bug fix and improvements
44c2c6d
-
update
257183f
-
fix
01ede16
-
update
3ef3c58 -
update
fa470f3 -
update
452ae9b -
Fix Chinese character encoding in tooltip display
Fixed UTF-8 decoding issue where Chinese characters were displayed as garbled text (e.g., 预计 instead of 预计) in the top-10 predictions tooltip. Root cause: JavaScript's atob() function only supports Latin-1 encoding, not UTF-8. When decoding base64-encoded JSON containing Chinese characters, each UTF-8 byte was incorrectly interpreted as a separate Latin-1 character. Solution: Implemented proper UTF-8 decoding by: 1. Using atob() to decode base64 to binary string 2. Converting binary string to Uint8Array 3. Using TextDecoder('utf-8') to correctly decode UTF-8 bytes Co-Authored-By: Claude Sonnet 4.5 <noreply@anthropic.com>04ab453 -
Simplify UI labels and remove redundant instructions
Changes: 1. Updated header labels in HTML: - "Compression A" → "RWKV Compression" - "Compression B" → "Qwen Compression" 2. Removed redundant instruction lines from app.py: - Removed color legend explanation (already in visualization) - Removed hover instruction (self-explanatory) - Kept only the main comparison description This makes the UI cleaner and less cluttered. Co-Authored-By: Claude Sonnet 4.5 <noreply@anthropic.com>6bbbdc0 -
Replace generic labels with specific model names
Changes in legend: - "Model A better than avg" → "RWKV better than avg" - "Model A worse than avg" → "RWKV worse than avg" Changes in tooltip: - "Compression A" → "RWKV Compression Rate" - "Compression B" → "Qwen Compression Rate" - "Model A Top10" → "RWKV Top10" - "Model B Top10" → "Qwen Top10" This makes the UI more intuitive by using actual model names instead of generic Model A/B labels throughout. Co-Authored-By: Claude Sonnet 4.5 <noreply@anthropic.com>6a2a99c -
Update tooltip labels to show model names directly
Changes: - Changed "Model A:" to "RWKV:" in tooltip - Changed "Model B:" to "Qwen:" in tooltip - Updated CSS classes from .qwen/.rwkv to .model-a/.model-b - Maintained color scheme: RWKV (#fcd34d), Qwen (#7dd3fc) This makes the tooltip more intuitive by showing actual model names instead of generic "Model A/B" labels. Co-Authored-By: Claude Sonnet 4.5 <noreply@anthropic.com>41790f6 -
Update legend text for better accuracy
Changed legend descriptions to be more precise: - "Model A better" → "Model A better than avg" - "= Avg delta" → "Equal to avg" - "Model B better" → "Model A worse than avg" This clarifies that colors represent Model A's performance relative to the average delta, not a direct comparison between models. Co-Authored-By: Claude Sonnet 4.5 <noreply@anthropic.com>6bb528b -
Improve UI and fix multi-token display in tooltips
- Increase max text length from 4000 to 8192 characters - Rename model labels to use generic Model A/B instead of hardcoded names - Fix tooltip to show all tokens in a byte range (not just one) - Remove "UncheatableEval" title from HTML output - Update data attributes from data-qwen/data-rwkv to data-model-a/data-model-b Co-Authored-By: Claude Sonnet 4.5 <noreply@anthropic.com>6fcf271 -
Remove debug logging and swap Model A/B positions
Changes: - Removed all debug print statements from app.py and html_generator.py - Swapped model positions: RWKV7 is now Model A, Qwen3 is now Model B - Green now indicates RWKV7 performs better, Red indicates Qwen3 performs better Co-Authored-By: Claude Sonnet 4.5 <noreply@anthropic.com>fa6172d -
Fix JSON truncation by using base64 encoding
Replaced direct JSON embedding in HTML attributes with base64 encoding to avoid issues with special characters breaking HTML attribute boundaries. This completely eliminates the escaping problems that were causing the "Unexpected end of JSON input" error in Top 10 predictions. Changes: - Encode topk JSON data as base64 before embedding in data attributes - Decode base64 in JavaScript before parsing JSON - Update debug logging to show base64 encoding status Co-Authored-By: Claude Sonnet 4.5 <noreply@anthropic.com>f59198d -
Add debug logging for JSON escaping in HTML attributes
Added detailed debug output to track how JSON is escaped before being embedded in HTML data attributes. This will help identify where the truncation is occurring. Co-Authored-By: Claude Sonnet 4.5 <noreply@anthropic.com>491ce2b -
Fix JSON truncation in HTML attributes for Top 10 predictions
Enhanced escape_for_attr function to properly escape all special characters that could break HTML attributes, including newlines, tabs, and quotes. This fixes the "Unexpected end of JSON input" error in Top 10 predictions. Co-Authored-By: Claude Sonnet 4.5 <noreply@anthropic.com>56292f7 -
Add detailed error logging for Top 10 predictions debugging
- Log error details to browser console - Display error message in tooltip - Output topkJson content for inspection Co-Authored-By: Claude Sonnet 4.5 <noreply@anthropic.com>69f914e -
Fix token display to show actual token IDs instead of positions
- Show actual model token IDs from topk_predictions - Decode token IDs using the correct tokenizer - Display format: [token_id] 'decoded_text' - Add fallback to display tokenizer info if token ID unavailable Co-Authored-By: Claude Sonnet 4.5 <noreply@anthropic.com>8629056 -
Optimize debug logging to reduce output volume
- Only print detailed debug info for first token - Add token_count variable to track processing - Print sample JSON output for verification - Reduce console spam while maintaining debugging capability Co-Authored-By: Claude Sonnet 4.5 <noreply@anthropic.com>24f1b39 -
Add comprehensive debug logging for Top 10 predictions
- Add debug prints in app.py to track evaluation results - Log tokenizer types and top5_predictions data structure - Add detailed logging in html_generator.py for token mapping - Track topk JSON generation process with error details - Print pred structure and decoding attempts for debugging Co-Authored-By: Claude Sonnet 4.5 <noreply@anthropic.com>951e8dd -
Fix Top 10 predictions display error
- Improve decode_token function with better error handling - Add detailed logging for debugging token decoding issues - Enhance build_byte_to_token_map with more informative error messages - Add docstring to build_byte_to_token_map function Co-Authored-By: Claude Sonnet 4.5 <noreply@anthropic.com>52ba00f -
Initial commit: UncheatableEval LossLens visualization
- Compare Qwen3-1.7B-Base vs RWKV7-G1C-1.5B byte-level predictions - Interactive HTML visualization with hover tooltips - Support both CPU and GPU execution - Gradio web interface for HuggingFace Spaces Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>9afeeeb