⚠️ Deprecated — preliminary (Phase 2 · The Stack v1)
This adapter is the original March-2026 hackathon (Phase 2) model, trained on bigcode/the-stack (v1, non-dedup). It is superseded by the paper's Phase 3 adapter, which was re-trained from scratch on the cleaner bigcode/the-stack-v2-dedup corpus. For paper-grade use, load the Phase 3 adapter from the umbrella repo:
PeftModel.from_pretrained(base_model, "legesher/language-decoded-lora", subfolder="tiny-aya-base/condition-3-zh-5k-native-code-seed42")
This repo is kept for reproducibility of the preliminary results only — do not cite it for the paper. It was renamed from legesher/language-decoded-lora-condition-3-zh-5k; the old URL continues to resolve via a Hugging Face redirect.
Blended dataset of 3,486 native Chinese code files + 1,514 transpiled Python (5k subset), assembled in Phase 2. Tests whether diverse native-language code adds value beyond reserved-word swapping.
Part of the Language Decoded project (Cohere's Tiny Aya Expedition).
For the full adapter inventory across both phases, see the Language Decoded LoRA hub and its MANIFEST.md.
Training Data
legesher/language-decoded-data / phase-2-the-stack-v1-condition-3-zh-5k — the Phase 2 / The Stack v1 config.
Usage
from transformers import AutoModelForCausalLM, AutoTokenizer
from peft import PeftModel
base_model = AutoModelForCausalLM.from_pretrained("CohereLabs/tiny-aya-base")
tokenizer = AutoTokenizer.from_pretrained("CohereLabs/tiny-aya-base")
model = PeftModel.from_pretrained(base_model, "legesher/language-decoded-lora-phase-2-the-stack-v1-condition-3-zh-5k")
Citation
@misc{language-decoded-2026,
title={Language, Decoded: Exploring the Impact of Fine-Tuning a Multilingual Model on Native-Language Code},
author={Madison Edgar and Saad Ahmed Bazaz and Tom Sherborne and Rashik Shahjahan and Khojasteh Mirza and Sarah Jawaid and Rafay Mustafa and Sohaib Ahmed Bazaz},
year={2026},
publisher={Hugging Face},
url={https://huggingface.co/legesher/language-decoded-lora}
}
License
Apache 2.0