ministral3-14b-20k-lora-ddp-20260826

Ministral 3 14B Instruct 20K LoRA

This repository contains a PEFT/LoRA adapter fine-tuned from mistralai/Ministral-3-14B-Instruct-2512-BF16.

Training snapshot

• Architecture: Mistral3ForConditionalGeneration
• Training rows: 20,000
• Maximum sequence length: 1,024
• Optimizer steps: 1,250
• Training topology: two-node DDP
• Final artifact type: LoRA adapter
• Source run date: 2026-08-26
• Repository visibility at upload: private

Loading

The base model is not duplicated in this repository. Load the base model separately and then attach this PEFT adapter.

from transformers import AutoModelForImageTextToText
from peft import PeftModel

base_model_id = "mistralai/Ministral-3-14B-Instruct-2512-BF16"
adapter_id = "koreallmdev/ministral3-14b-20k-lora-ddp-20260826"

base = AutoModelForImageTextToText.from_pretrained(
    base_model_id,
    torch_dtype="auto",
    device_map="auto",
)
model = PeftModel.from_pretrained(base, adapter_id)

Validation status

The training run completed and the final adapter was saved successfully. At the time this upload package was prepared, the adapter had not yet been recorded as a promoted STABLE release and a post-training quality benchmark was not included in the upload metadata.

Do not interpret this repository as a benchmark claim unless evaluation results are added later.

Notes

Only the final adapter artifact is uploaded by this workflow. Intermediate training checkpoints and local training logs are intentionally excluded.