kuza-gemma-4-e2b

Kuza Gemma 4 E2B

Full training-run archive for Kuza (East Africa agricultural assistant), fine-tuned from unsloth/gemma-4-E2B-it-qat-q4_0-unquantized. Weights, logs, checkpoints, GGUFs, and provenance are stored with the same layout as $KUZA_WORK_DIR/kuza-gemma-4-e2b/.

This derivative is subject to the Gemma license.

Training mix

• 100% English train from kuzaai/kuza_sft_english
• 50% Swahili from kuzaai/kuza_sft_swahili
• 8% HuggingFaceH4/no_robots
• 5% adversarial from kuzaai/kuza_sft_adversarial
• 3% multiturn from kuzaai/kuza_sft_multiturn
• 10% code-switching (Sw-En) from kuzaai/kuza_sft_code_switch
• 6% passage-grounded QA from kuzaai/kuza_sft_grounding

LoRA: RsLoRA r=32, alpha=64, QAT int4. Sequence length 2048, 2 epochs, LR 0.0001. Thinking is off. GGUFs are text-only (PLE kept; vision/audio dropped).

Preference alignment (DPO)

adapter/ holds the SFT+DPO policy (promoted by 02b_dpo.py). Preference pairs: 866 train / 97 validation from kuzaai/kuza_dpo_preference. DPO: beta 0.1, RPO alpha 1.0, LR 5e-06, 1 epoch(s), loss sigmoid. The pristine SFT adapter is preserved in adapter-sft/; the pre-promotion DPO output is preserved in adapter-dpo/.

The preference dataset kuzaai/kuza_dpo_preference is maintained and uploaded separately by the repo owner; it is never pushed by the training pipeline.

Files

• adapter/ — PEFT policy adapter (SFT+DPO), tokenizer, metrics and manifests
• adapter-sft/ — frozen pristine SFT adapter (pre-alignment snapshot)
• adapter-dpo/ — DPO output as trained (pre-promotion copy)
• training/ — Trainer checkpoints including checkpoint-best
• merged_bf16/ — text-only merged Hugging Face BF16 weights
• reference/ — text-only BF16 GGUF (kuza-bf16.gguf) and smoke log
• reference_qat/ — QAT lattice-exact Q4_0 export (kuza-qat-q4_0-lattice.gguf)
• imatrix/ — calibration corpus, eval corpus, imatrix, logs
• quants/ — quantized GGUF candidates:
• iq4_nl/ (iq4_nl, from reference)
• iq4_xs/ (iq4_xs, from reference)
• q4_0_imatrix_pure/ (q4_0, from reference)
• q4_0_qat_export/ (q4_0, from reference_qat)
• q4_k_m_default/ (q4_k_m, from reference)
• q4_k_m_ud_style/ (q4_k_m, from reference)
• q8_0_ceiling/ (q8_0, from reference)
• ud_q4_k_xl/ (q4_k_m, from reference)
• screen/ — GPU KLD diagnostic, hidden-set scores, and results.json winner
• provenance/ — copied adapter metrics, recipes, screen JSON
• upload_manifest.json — path, size, and sha256 for every uploaded file

Download

huggingface-cli download kuzaai/kuza-gemma-4-e2b --local-dir ./kuza-gemma-4-e2b

screen/results.json applies the quality gates (safety, language, same-top-p, KLD ceiling) first, then ranks gate-passers within KLD noise by hidden-set accuracy, CPU generation t/s, GGUF size, and peak RSS. CPU numbers are the deployment signal for the edge target; GPU t/s and KLD are diagnostics.