Spaces:
Sleeping
Sleeping
Commit History
-
fix bug in prompt preview display
41ee8bf -
add prompt preview for debug
5fc0117 -
fix: prevent self-talking issue by using tokenizer chat_template formatting
- Updated prompt construction to use `tokenizer.apply_chat_template()` for chat-tuned models - Ensured backward compatibility with non-chat models by falling back to manual prompt formatting - Resolves issue where models would echo both user and assistant messages due to incorrect flat-text prompts960db60
-
bugfix to Error: "str" object has no attribute "pad_token_id"
889f080 -
add taiwan elm 1.1b & 270m instruct
c8399e3
-
add type in qwen3 0.6b repo id
76d4d60 -
add qwen3
fe395ab
-
Add Smollm2 360m instruct fine-tuned on TaiwanChat
7308211
-
keep debug message
37f7787 -
add debug to show web resarch result
a2f07a4 -
give 1 second for web search to grab data
9ad3ffd -
inject web search result if web search enabled
bc257ff -
refactor(app): improve streaming, background search, dtype fallback, and cleanup :contentReference[oaicite:0]{index=0}
- Add cancel_event checks in the streamer loop to enable true cancellation during response generation. :contentReference[oaicite:1]{index=1} - Launch DuckDuckGo web search in a background thread to prevent blocking the streaming pipeline. :contentReference[oaicite:2]{index=2} - Implement dtype fallback (bfloat16 → float16 → float32) for broader hardware compatibility. :contentReference[oaicite:3]{index=3} - Suppress repeated debug messages after the first token to avoid UI flooding. :contentReference[oaicite:4]{index=4} - Remove unused imports and streamline load_pipeline caching logic for cleaner code. :contentReference[oaicite:5]{index=5}293686e -
bugfixc: not using pipeline for response generation
939895d -
Add original SmolLM2 135M Instruct for comparaison
423dc1a -
Add SmolLM2-135M-Instruct-TaiwanChat
38fcc03 -
Add SmolLM2-135M TaiwanChat
0d642b7
-
default to gemma-3-4b
88a6a62 -
model repo_id typo fix
89372fa -
enable web search by default
6235e63 -
remove tinyllama which has bad response quality
a22cf42 -
make streaming response
5ea073d -
apply history flatten before it goint to prompt
ef361b0 -
better management on system prompt
5f6306a -
usue chat pipeline instead of model and tokenizer individually
ac8e9cc -
bugfix to padding-related issues
f248fec -
add attention mask
b6b3940 -
Clean model description
4731160 -
pin torch to 2.4.0
4c6b4c5 -
enable zerogpu with decorator
c5b2897 -
Apply ZeroGPU
d181b45 -
switch to gradio version for stability reason
a703203
-
add taiwan tinyllama
794ee70 -
Improve responsiveness by asynchronously retrieving web search context
- Run web search lookup in a background thread to decouple it from prompt construction. - Add a 2-second timeout using thread.join(timeout=2) to prevent delaying response generation. - Fallback to using only the system prompt and user query if web search results aren't retrieved in time. - Update sidebar UI to show the fetched context or a "No context found" message. - Retain current streaming and chat history functionalities for model inference.acda3f1 -
add llama 3.2 taiwan 3b
b1544e2 -
add minicpm3 4b
f5c0811 -
increase xt length to max
629495e -
remove all moe
fafc8cb -
remove qwen 1.5 moe
6735035 -
adjust title style
e9559bd -
use another version of qwen 1.5 moe
96e60d6 -
add Qwen1.5-MoE
e17afaf -
Qwen2.5-MOE-6x1.5B
5eca666 -
remove under 3b models
617be26 -
Add model caching
d33dfcd -
UI/UX Improvement
eb215ff -
reset timeout timer once a new token is generated
35943b1
-
open web search settgins to user
c9fd924 -
add 2 more models
f7a541f -
apply new settings on duckduck search
d9421eb