kuza-qwen-3.5-4b

Kuza Qwen 3.5-4B

Full training-run archive for Kuza (East Africa agricultural assistant), fine-tuned from unsloth/Qwen3.5-4B. Weights, logs, checkpoints, GGUFs, and provenance are stored with the same layout as $KUZA_WORK_DIR/kuza-qwen-3.5-4b/.

This derivative is subject to the Apache 2.0 license of the Qwen base model.

Training mix

• 100% English train from kuzaai/kuza_sft_english
• 35% Swahili from kuzaai/kuza_sft_swahili
• 8% HuggingFaceH4/no_robots
• 5% adversarial from kuzaai/kuza_sft_adversarial
• all multiturn from kuzaai/kuza_sft_multiturn

LoRA: RsLoRA r=32, alpha=64, no QAT. Sequence length 1024, 2 epochs, LR 0.0002. Thinking is off (--reasoning off when llama.cpp supports it). GGUFs are text-only (MTP/nextn kept; vision/audio dropped). ssm_out is Q6_K on every candidate.

Files

• adapter/ — PEFT adapter, tokenizer, SFT metrics and manifests
• 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
• imatrix/ — calibration corpus, eval corpus, imatrix, logs
• quants/ — quantized GGUF candidates:
• q4_k_m_imatrix/kuza-qwen-q4_k_m.gguf (q4_k_m)
• q4_k_xl_ssm/kuza-qwen-q4_k_xl.gguf (q4_k_m)
• q4_k_s_ssm/kuza-qwen-q4_k_s.gguf (q4_k_s)
• 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-qwen-3.5-4b --local-dir ./kuza-qwen-3.5-4b

screen/results.json ranks by hidden-set accuracy, then GGUF size, then GPU TPS. Do not treat GPU TPS as an ADTC laptop measurement.