kuclab-hertz-0.5

KucLab Hertz 0.5

A Czech/English STEM + programming + web-dev assistant built by KucLab on top of google/gemma-4-12B-it, with a more direct, human personality (dry/dark humor allowed, minimal reflexive hedging) and a stronger focus on clean, non-generic code output.

What this is

Hertz 0.5 is a LoRA fine-tune (r=8, merged into the base weights) trained on a self-distilled corpus. Training data was generated by prompting Qwen3.8-27B (Alibaba/Tongyi, Apache 2.0) as a data-generation teacher — it never trains itself, only produces answers used to fine-tune the actual model that ships here.

• Base: google/gemma-4-12B-it (11.95B params, Apache 2.0)
• Method: QLoRA, r=8 / alpha=16, merged to bf16 then quantized
• Context: 262144 tokens natively (inherited from base). Deployed with a default num_ctx of 65536.
• Training data: 1036 rows — 285 newly self-distilled this round (STEM concepts/terms, personality/humor, web-dev components, agentic step-by-step planning), plus 736 rows reused as-is from Hertz 0.4's corpus.
• New this release: direct/witty personality (dry or dark humor allowed when it fits, fewer reflexive "as an AI I cannot..." hedges on ordinary topics — genuinely harmful requests are still refused), an explicit anti-"AI slop" instruction for generated web code (no generic gradients, no glassmorphism-for-its-own-sake, no cliché marketing copy), and a small set of agentic step-by-step planning examples.
• Format available: GGUF (q4_k_m, ~7.4GB) for llama.cpp/Ollama, plus the raw LoRA adapter.

Quickstart (Ollama)

Important: ollama pull hf.co/... alone does NOT apply this model's system prompt (identity + personality) — Ollama only fetches the raw GGUF from Hugging Face, it does not read a repo's Modelfile. Without the system prompt, the model falls back to identifying as a generic Gemma model. Use ollama create with the Modelfile below instead — it pulls the weights AND applies the system prompt in one step:

curl -O https://huggingface.co/KucLab/kuclab-hertz-0.5/resolve/main/Modelfile
ollama create kuclab-hertz-0.5 -f Modelfile
ollama run kuclab-hertz-0.5

(The Modelfile's FROM line points at hf.co/KucLab/kuclab-hertz-0.5:Q4_K_M, so this pulls the same GGUF automatically — no separate download needed.)

Example

Q: Jsi cenzurovaný?

A: Cenzura v odborném smyslu neexistuje – odmítnu jen to, co je objektivně škodlivé nebo nelegální.

Q: Napiš jednoduchou pricing table komponentu v HTML/CSS, tři tarify, moderní ale ne AI slop.

A: (clean semantic HTML/CSS, no gradients, no glassmorphism, plain honest copy)

Benchmarks — full, honest numbers, including a real regression we found

Both models measured with the same prompts, same grading code, same Ollama Q4_K_M quantization, cold. Important process note: our first benchmark pass used a 300-second per-question timeout that turned out to be too short once we raised the token budget — the client would discard the entire response as <ERR TimeoutError> if the model was still thinking past 300s, which happened more often at a bigger token budget, not less. That first pass showed Hertz 0.5 at 48.8%, which was an artifact, not a real score. We fixed the timeout (1200s) and reran both models with the identical corrected methodology. The numbers below are from that clean rerun.

MMLU-Pro STEM (240 held-out questions, this project's own curated subset — not the full public MMLU-Pro set, so this is not directly comparable to Google's official published MMLU-Pro number for this base model)

This is a real, measured regression, not a benchmark artifact this time — both models were run with the identical, corrected methodology, and the drop is consistent across all four subjects. Our read: this round's fine-tune added only 285 genuinely new rows (the rest reused from 0.4) and focused them on personality, web-dev, and agentic planning rather than STEM depth. A small-rank LoRA adapter (r=8, ~0.2% of parameters) trained this way appears to trade a real slice of raw STEM multiple-choice accuracy for the stylistic/personality shift. We're not aware of a way to have both without more, and more STEM-focused, training data — which is planned for the next release.

Czech terminology benchmark (206 held-out CS↔EN scientific terms)

The best Czech-terminology result in this project's history. (Hertz 0.4 never had this benchmark run — not skipped on purpose, just not gotten to before 0.5 started.)

For context: Google's own published MMLU-Pro number for google/gemma-4-12B-it is 77.2% (full ~12k-question test, official eval harness) — close to our own base-model measurement (75.8%) on our smaller STEM-only subset, which is a useful sanity check that our subset isn't wildly out of line with the official full-set number, even though the two aren't a strict apples-to-apples comparison.

Honest status

• ✅ Coherent, correct Czech and English output; personality genuinely shifted toward direct/witty as intended (spot-checked)
• ✅ Correctly identifies as a KucLab model, not as "Gemma" (LoRA + system prompt)
• ✅ Web-dev output avoids the generic-AI-slop patterns it was trained against (spot-checked, not benchmarked numerically — no such benchmark exists yet)
• ✅ Czech terminology: best result in project history (75.7%, +10.2pp over 0.3)
• ⚠️ MMLU-Pro STEM is measurably lower than the un-fine-tuned base model (68.8% vs 75.8%, −7.0pp) — confirmed with corrected, identical methodology on both models. This is disclosed, not hidden.
• ⏳ Tool-calling fine-tuning — base supports it natively, this fine-tune didn't add tool-use training examples
• ⏳ No uncensoring/decensoring pass in the sense of removing real safety boundaries — the personality shift is about dropping unnecessary hedging on ordinary topics, not about disabling refusals for genuinely harmful requests
• ⏳ LoRA adapter upload pending a local file-permission fix — will be added to this repo shortly after the GGUF

If you're deciding whether this fits your use case: it's a friendlier, funnier, more web-dev-capable "gemma-4-12B-it, nudged toward KucLab identity and Czech fluency," at a real, disclosed cost to STEM multiple-choice accuracy versus the base model. If pure STEM benchmark performance is your priority, the base model currently scores higher on our own measurement.

License

Apache 2.0, inherited from google/gemma-4-12B-it (per Google's official Hugging Face listing). Qwen3.8-27B (used only to generate training data, never trained or redistributed here) is separately licensed under Apache 2.0 by Alibaba/Tongyi.

Credits

• Base model: google/gemma-4-12B-it (Google, Apache 2.0)
• Training-data teacher: Qwen/Qwen3.8-27B (Alibaba/Tongyi, Apache 2.0)
• Fine-tuning, dataset construction, and packaging: KucLab