Small corpus (3,285 pairs) — limits generalization
Literary domain only (19th century books) — no modern vocabulary like "tekoäly" (AI) or "koneoppiminen" (ML)
Repetition loops observed on out-of-domain prompts
v2 planned with mixed-domain corpus (Opus-100 + WikiMatrix + Tatoeba, ~65k pairs)

Usage

python

from peft import PeftModel
from transformers import AutoModelForCausalLM, AutoTokenizer

base = AutoModelForCausalLM.from_pretrained("mistralai/Mistral-7B-Instruct-v0.2")
model = PeftModel.from_pretrained(base, "indrang21/mistral-7b-fi-en-qlora-v1")
tokenizer = AutoTokenizer.from_pretrained("indrang21/mistral-7b-fi-en-qlora-v1")

inputs = tokenizer("[FI] Tervetuloa. [EN]", return_tensors="pt")
outputs = model.generate(**inputs, max_new_tokens=50)
print(tokenizer.decode(outputs[0]))

Framework

Full code: https://github.com/indrang21/Suomi-RAG-er-bilingual-embedding-retrieval

indrani191919/mistral-7b-fi-en-qlora-v1

Mistral 7B Finnish-English QLoRA (v1)

Model Details

Training Config

Results

Limitations (Honest)

Usage

Framework

indrani191919/
mistral-7b-fi-en-qlora-v1