Dental AI research SLM — continued domain knowledge from Polytechnique publications. Trained on 2448 documents (~1677519 tokens) from the dental.ai.research domain.
Small model — capabilities are proportionally limited
Domain specialization may cause out-of-domain degradation vs baseline
Not a general-purpose assistant; best at dental.ai.research-specific tasks
Recommended sampling settings
This is a 0.5B model fine-tuned on a narrow domain (~1.7M tokens). Without
the parameters below, it can collapse into repetition or emit garbled GGUF
detokenize artifacts (e.g. literal tti, ttiassistant). Always use:
param
value
reason
temperature
0.5
0.7 is too hot for a 0.5B base on narrow corpus
top_p
0.9
constrains tail
top_k
40
constrains tail
repeat_penalty
1.18
kills paragraph loops (the main failure mode)
repeat_last_n
256
window for the penalty
max_tokens
320
keep responses short
LM Studio: open the right-hand "Advanced configuration" panel and set
the values above. Also add these stop strings under "Stop strings":
<|im_end|>, <|im_start|>, <|endoftext|>, ttiuser, ttiassistant.
Ollama: the published Modelfile has these defaults baked
in — just run ollama create ... -f Modelfile.
llama.cpp / llama-cpp-python: see the Space app.py
for a known-working configuration.
How to Use
Download the GGUF for local inference (LM Studio, llama.cpp, text-generation-webui)
Quantizations available:
gguf/model-Q4_K_M.gguf — ~397.8 MB — best size/quality trade-off
