MiniCPM5-1B-ru-lora-v3

MiniCPM5-1B — Russian LoRA (v3, experimental)

A LoRA adapter (rank 32) for openbmb/MiniCPM5-1B, aimed at improving practical Russian (customer-support style, everyday instructions, translation) while keeping the base's coding ability. Trained on Vikhrmodels/GrandMaster-PRO-MAX (Russian + English, with chain-of-thought) plus an English replay slice, on an AMD Strix Halo iGPU (ROCm).

Honest evaluation (base vs +LoRA, thinking OFF, 3-run avg on local batteries)

What it does well: practical Russian gets noticeably more fluent and coherent — customer-support replies stop mixing English/Chinese and become natural Russian; RU→EN translation is more accurate.

The tradeoff: sharp reasoning and strict instruction-following drop somewhat — an inherent capacity limit of a 1B model (you can't add a language without spending some of its other skills). For strong Russian without the tradeoff, consider a Russian-native model or a larger base.

Use thinking OFF (enable_thinking=false) — the adapter was trained for the non-thinking format.

Usage

Transformers (PEFT):

from transformers import AutoModelForCausalLM, AutoTokenizer
from peft import PeftModel
base = AutoModelForCausalLM.from_pretrained("openbmb/MiniCPM5-1B", trust_remote_code=True)
model = PeftModel.from_pretrained(base, "vahpetr/MiniCPM5-1B-ru-lora-v3")
tok = AutoTokenizer.from_pretrained("openbmb/MiniCPM5-1B", trust_remote_code=True)

llama.cpp (GGUF adapter MiniCPM5-1B-ru-lora-v3.gguf is included):

llama-server -m MiniCPM5-1B-Q8_0.gguf --lora MiniCPM5-1B-ru-lora-v3.gguf --jinja

(A merged, ready-to-run version is at vahpetr/MiniCPM5-1B-ru-v3.)

Attribution & license

• Base: MiniCPM5-1B (OpenBMB) — Apache-2.0
• Data: GrandMaster-PRO-MAX (Vikhr) — Apache-2.0
• This adapter — Apache-2.0