luispoveda93/MiniCPM5-2B-catalan-chat-v2-GGUF

MiniCPM5-2B-catalan-chat-v2 — GGUF

GGUF conversions of [luispoveda93/MiniCPM5-2B-catalan-chat-v2](https://huggingface.co/luispoveda93/MiniCPM5-2B-catalan-chat-v2) — round 2: a LoRA fine-tune of luispoveda93/MiniCPM5-2B-catalan-chat on 23.4K multi-turn Catalan conversations from BSC-LT/ALIA-2606-SFT (CC-BY-4.0), focused on chat behaviour.

Converted with llama.cpp convert_hf_to_gguf.py (fp16) and quantized with llama-quantize (release b10067). The chat template is embedded in the GGUF metadata (tokenizer.chat_template).

Available quants

Usage

llama.cpp

llama-server -m MiniCPM5-2B-catalan-chat-v2-Q4_K_M.gguf --port 8080

Ollama

FROM MiniCPM5-2B-catalan-chat-v2-Q4_K_M.gguf
SYSTEM "Ets un assistent conversacional que respon sempre en català."

ollama create minicpm5-catalan-v2 -f Modelfile
ollama run minicpm5-catalan-v2

Details

• —Base model (round 2): luispoveda93/MiniCPM5-2B-catalan-chat — round-1 LoRA on openbmb/MiniCPM5-2B (2.5B params, Llama architecture, vocab 130,560)
• —Round-2 training: 23.4K multi-turn Catalan conversations from BSC-LT/ALIA-2606-SFT (multi-turn augmentation, instruction-following, mentor-ca, dolly-ca, CoQCat, identity, system-prompt multi-turn), LoRA r=32/α=64, lr 1e-4, 1 epoch, ~326 steps; final eval loss 1.383, token accuracy 0.704 (metrics)
• —License: base model Apache-2.0; round-2 data CC-BY-4.0; round-1 lineage contains InstruCAT (CC-BY-NC-ND-4.0, non-commercial)
• —Limitations: 2.5B params; chat behaviour focus — task-QA skills come from round 1's single-turn data