README.md

---
license: apache-2.0
base_model: google/gemma-4-12B-it
language:
- cs
- en
tags:
- gemma4
- czech
- stem
- physics
- chemistry
- biology
- mathematics
- programming
- lora
- self-distillation
pipeline_tag: text-generation
---

# KucLab Hertz 0.4

A Czech/English STEM + programming assistant built by [KucLab](https://kuclab.org) on top of **google/gemma-4-12B-it**, fine-tuned toward physics, chemistry, biology, mathematics and programming in Czech, while keeping the base model's native 256k-token context and tool-calling.

## What this is

Hertz 0.4 is a **LoRA fine-tune** (r=8, merged into the base weights) trained on a self-distilled corpus of Czech/English STEM concepts, worked problems, programming exercises, and formatting examples. 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 — with a system instruction asking for direct, confident answers over hedged ones.

- **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 (no context-extension hack needed — unlike Hertz 0.3's Qwen2.5-14B base, this model ships long context out of the box). Deployed here with a default `num_ctx` of 65536; the model can go higher, this is a practical default, not a hard limit.
- **Training data:** 744 rows — 497 newly self-distilled (STEM concepts/terms/problems, a new programming category, formatting), plus 234 rows reused as-is from Hertz 0.3's corpus.
- **Format available:** GGUF (q4_k_m, ~7.4GB) for `llama.cpp`/Ollama, plus the raw LoRA adapter.

Gemma-4 has its own native thinking mechanism (channel-based, gated by an `enable_thinking` flag). Training data deliberately does **not** replicate that token syntax — `output` is plain text with visible step-by-step reasoning, same choice as Hertz 0.3. This doesn't disable native thinking mode, it just means the fine-tune doesn't touch it either way.

## 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:

```bash
curl -O https://huggingface.co/KucLab/kuclab-hertz-0.4/resolve/main/Modelfile
ollama create kuclab-hertz-0.4 -f Modelfile
ollama run kuclab-hertz-0.4
```

(The Modelfile's `FROM` line points at `hf.co/KucLab/kuclab-hertz-0.4:Q4_K_M`, so this pulls the same GGUF automatically — no separate download needed.)

## Example

> **Q:** Kolik je 17 × 23? Ukaž postup.
>
> **A:** Výsledek: 391 — pak krok za krokem 17×(20+3) = 340+51 = 391.

> **Q:** Napiš v Pythonu funkci pro binární vyhledávání.
>
> **A:** *(correct, idiomatic implementation with type hints and a docstring)*

## Benchmarks — and an honest complication

Measured against the same base model (`google/gemma-4-12B-it`, Q4_K_M via Ollama), same prompts, same grading code, both models cold.

**MMLU-Pro STEM** (240 held-out questions, this project's own curated subset)

| | base | Hertz 0.4 | Δ |
|---|---|---|---|
| Biology | 86.7% | 76.7% | −10.0pp |
| Chemistry | 55.0% | 45.0% | −10.0pp |
| Math | 85.0% | 76.7% | −8.3pp |
| Physics | 63.3% | 56.7% | −6.6pp |
| **Total** | **72.5%** | **63.7%** | **−8.8pp** |

**We do not trust this number at face value, and you shouldn't either.** Both models — base and fine-tuned — tend to work through every multiple-choice option individually with detailed reasoning before giving a final answer, and a meaningful fraction of responses got cut off by our grading harness's token budget before reaching the final answer letter (18% unparsed for base, 23% for Hertz 0.4). The drop is uniform across all four subjects, which points to a harness/verbosity artifact rather than genuine domain-knowledge loss, but we have not yet re-run this with a larger token budget to confirm. Until we do, **treat this MMLU-Pro number as unverified and likely pessimistic for both models** — we're publishing it anyway because we'd rather show a confusing number with an honest explanation than hide it.

Czech terminology benchmark: not yet run for this release.

## Honest status

- ✅ Coherent, correct Czech and English output (spot-checked: arithmetic, physics, code, terminology — one wrong term found: "smělná odchylka" instead of "směrodatná odchylka" for standard deviation)
- ✅ Correctly identifies as a KucLab model, not as "Gemma" (LoRA + system prompt)
- ✅ 256k native context (inherited from base, not independently long-context-tested by us yet)
- ⚠️ MMLU-Pro STEM comparison run but confounded by a benchmark-harness token-budget issue (see above) — re-run pending
- ⏳ Czech terminology benchmark not yet run for this release
- ⏳ Tool-calling fine-tuning — base supports it natively, this fine-tune didn't add tool-use training examples
- ⏳ No uncensoring/decensoring pass — safety behavior inherited from the base model as-is

If you're deciding whether this fits your use case: treat it as "gemma-4-12B-it, nudged toward Czech STEM/programming fluency and KucLab identity," with an open question mark on general STEM benchmark performance until we re-verify it.

## 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](https://huggingface.co/google/gemma-4-12B-it) (Google, Apache 2.0)
- Training-data teacher: [Qwen/Qwen3.8-27B](https://huggingface.co/Qwen/Qwen3.8-27B) (Alibaba/Tongyi, Apache 2.0)
- Fine-tuning, dataset construction, and packaging: [KucLab](https://kuclab.org)