peerbench/gemma4-german-sft-corpus

Gemma-4-E4B German SFT Corpus — 4 controlled variants

Curated, native-heavy German supervised-fine-tuning (SFT) corpus, built to improve the general German skill of unsloth/gemma-4-E4B-it via LoRA — NOT to target any single benchmark. The EuroEval-ported German benchmarks (scala_de, sb10k_de, include_de, mmlu_prox_de, germeval_de, germanquad_de, …) are used only as honest thermometers, never as training signal — no benchmark train/test split is mixed in, deliberately, to avoid Goodharting.

Four variants share one instruction/response backbone and differ only along three controlled axes, so that one LoRA per variant isolates each axis's effect.

Variant	MCQ target	Answer balance	Replay	Rows	Sources	Isolates
v1	reasoning	balanced	✅ ~8%	27 996	12	full recipe
v2	terse	balanced	❌	25 756	11	terse-vs-reasoning
v3	reasoning	balanced	❌	25 756	11	replay effect (vs v1)
v4	reasoning	unbalanced	✅ ~8%	27 996	12	balance effect (vs v1)

Designed contrasts (hold everything else equal):

• v1 ↔ v3 → replay effect (raw-German continuation as an anti-forgetting regularizer)
• v1 ↔ v4 → answer-balance effect (shuffling MCQ gold positions to kill the "default-to-A" bias)
• v2 ↔ v3 → terse-vs-reasoning MCQ target format

Format

ShareGPT-style JSON Lines, one conversation per line:

{
  "source": "maxidl/OpenOrca-gpt4-de",
  "conversations": [
    {"from": "human", "value": "..."},
    {"from": "gpt",   "value": "..."}
  ]
}

• Strict role alternation human → gpt → human → gpt …, starting at human, ending at gpt. Required by the Gemma-4 chat template, which rejects a system role and refuses consecutive same-role turns ("Conversation roles must alternate"). During the build, system turns were merged into the first user turn and consecutive same-role turns were collapsed.
• source records the upstream dataset, or a synthetic bucket: replay_de, m_arc_de, mmlu_aux_de.

Composition

Each variant = a shared bulk of 24 640 instruction/response rows (after dedup + per-source cap), + 1 116 MCQ rows (m_arc_de), + 2 240 replay rows for v1/v4. Per-source counts (v1 / v4; v2 / v3 are identical minus the 2 240 replay_de rows):

Rows	Source	Kind
3 485	FreedomIntelligence/alpaca-gpt4-deutsch	high-quality MT (GPT-4)
3 484	toroe/Soofi-Think-SFT-V2-firsthalf-DE	reasoning (math/science)
3 483	CausalLM/GPT-4-Self-Instruct-German	native German
3 482	flozi00 (/conversations)	native German
3 480	stefan-it/secret-german-reasoning	native German reasoning
3 477	maxidl/Capybara-de	high-quality MT
3 460	maxidl/OpenOrca-gpt4-de	high-quality MT (GPT-4)
2 240	replay_de — raw German text (Wikipedia), text-continuation (v1/v4 only)	replay
1 116	m_arc_de — alexandrainst/m_arc (de), balanced MCQ	MCQ
268	LeoLM/OpenSchnabeltier	native German
17	aya_de — CohereLabs/aya_dataset (de)	native German
4	mmlu_aux_de — MMLU auxiliary-train (de)	MCQ aux

Round-1 residue: aya_de (17) and mmlu_aux_de (4) are small because streaming missed most German aya rows and mmlu_aux collapsed on final-turn dedup. Retained for transparency; a round-2 rebuild will restore them.

Curation

• Native-heavy: native German prioritized; high-quality MT (DeepL'24 / GPT-4 translations) admitted; noisy MT excluded.
• Language filter: lingua German language-ID where available, with a stop-word + length heuristic fallback when lingua is absent.
• Per-source 20 % cap + round-robin mixing → no single source dominates. For the target of 28 000 the cap is 5 600 rows/source; in practice the 7 large sources each land at ~3 480 rows (well under the cap) because the round-robin draws evenly across them.
• Global dedup on the final gpt turn.
• Regularized approximate replay (replay_de, v1/v4 only): ~8.7 % raw-German Wikipedia text as human (passage start) → gpt (continuation), to counter catastrophic forgetting of base German fluency. Replay fraction = n · ratio / (1 − ratio).
• MCQ handling (m_arc_de): balanced variants shuffle the gold-answer position uniformly over A/B/C/D (kills positional / A-default bias); the unbalanced variant (v4) keeps original positions. terse target = "Die korrekte Antwort ist (X)."; reasoning target = short rationale then the letter.
• Explicitly excluded: slimorca_dedup_german_experimental (noisy MT), ultra_chat_de (verbosity).

Contamination

Independent of the held-out benchmarks (INCLUDE-de, EuroEval ScaLA / SB10k / GermEval / GermanQuad, MMLU-ProX / MMMLU-de). No benchmark test split is included; openai/MMMLU[DE] and LeoLM/MMLU_de (eval-only) were deliberately not used.

Intended use & trained models

LoRA SFT of unsloth/gemma-4-E4B-it: r128 / α256 · bf16 · lr 5e-5 (gentle) · effective batch 16 · 1 epoch · reasoning-off · seed 3407 · 1× H100. Models trained from these variants:

• peerbench/gemma-4-E4B-20260610190427-v1 — reasoning + balanced + replay
• peerbench/gemma-4-E4B-20260610190427-v2 — terse + balanced
• peerbench/gemma-4-E4B-20260610190427-v3 — reasoning + balanced, no replay
• peerbench/gemma-4-E4B-20260610190427-v4 — reasoning + unbalanced + replay

Provenance & licensing

Built by re-formatting and filtering the upstream datasets listed above; each retains its original license / terms — consult the source repositories before redistribution. Synthetic buckets (replay_de, m_arc_de, mmlu_aux_de) are transformations of publicly available German text / ARC / MMLU-auxiliary material. Provided as-is, for research; license: other reflects the mixed upstream terms.

Build

• Builder: build_gemma_sft_variants.py — --target 28000 --seed 3407.
• Built 2026-06-10 on the German-LLM GPU client host (gaa-mdk): ShareGPT normalization → language filter → per-source cap → dedup → strict-alternation pass → variant assembly.

---

Part of the peerbench German-LLM fine-tuning experiments.