EvidenceDesk

A support copilot that answers only from retrieved documentation, cites verbatim quotes, abstains when the evidence is missing, and can only propose a note that a human must approve. Generation runs on a fine-tuned, 4-bit Gemma 4 12B distilled from Gemma 4 26B-A4B and trained on a laptop.

Loading recorded runs…

Recorded runs

Evaluation

Cloudflare Llama 3.1 8BGemma 4 12B QAT Q4_0Fine-tuned Q4_0
Regression set, end-to-end pass (30 generations)0.8000.8670.900
Regression set p50 / p951.7 s / 3.1 s (cloud GPU)7.1 s / 22.3 s2.6 s / 6.1 s
Held-out products, pass rate (104, EN + TR)–0.9420.971
Held-out Turkish–0.9250.975
Held-out without grammar-constrained decoding–0.2690.904
Mean output tokens–244109

Gemma rows run with llama.cpp on an Apple M5 Pro. Held-out products and the regression product were never in the training data. Small samples; labels are not human-reviewed; a valid citation is not proof of semantic correctness.

Run it live

This page is a static Space, the only kind a free Hugging Face account can host, so it shows recorded runs. The interactive ZeroGPU app is in deploy/hf-space. Locally, the model runs with llama.cpp:

llama-server --hf-repo UmitVice/evidencedesk-gemma-4-12b-GGUF --hf-file evidencedesk-gemma-4-12b-Q4_0.gguf \
  --jinja -c 8192 --cache-ram 0 --port 8080
# then point EvidenceDesk at it:
GENERATION_PROVIDER=openai_compatible GENERATION_BASE_URL=http://127.0.0.1:8080/v1