Iman998/HexaBee-Multilingual-SFT

HexaBee · Multilingual Answering & Translation

Ask in one language. Answer in another.

Paper: Iman Barati, Mohammadmostafa Rostamkhani, and Mohammad Badzohreh, HexaBee-4B: Six-Language Translation with Multilingual Answering and Bidirectional Supervision, arXiv:2610.16361 (2026).

Overview · Quick start · Contents · Provenance · Limitations · Release family · Research code

Overview

The exact 22 final SFT exports used in phase 2: 19,458,332 answering records, including 5,759,806 cross-language QA records, plus 10,695,072 answer-text translations. The question language and requested answer language can differ; answering a question is distinct from translating its wording.

Property	Value
Artifact	HexaBee · Multilingual Answering & Translation
Scope	Phase 2 · 30,153,404 SFT records
Released rows	30,153,404
Format	Sharded Parquet + metadata
Release	Audited historical artifacts
Model family	HexaBee-4B / Gemma 3

Quick start

from datasets import load_dataset

ds = load_dataset("Iman998/HexaBee-Multilingual-SFT", "gqa_complete", split="train", streaming=True)
print(next(iter(ds)))

Contents

22 configurations named by neutral source group and task. system, instruct, output are original training strings. Component/task metadata and unique positional content hashes are added. Per-row language labels were stripped in the final original exports and are null here, rather than guessed.

release_manifest.json lists distributed files, sizes, and SHA-256. release_stats.json records measured artifact metadata. Data files use bounded Parquet shards; original strings and empty outputs are preserved in historical exports.

Available configurations

Configuration	Splits	Rows
gqa_complete	train	3,952,557
gqa_translate	train	2,481,368
qa_complete	train	7,240,185
qa_formal	train	1,540,222
qa_translate	train	5,554,592
title_complete	train	3,835,308
title_formal	train	646,570
title_informal	train	708,374
title_translate	train	2,659,112
gqa_balance	train	208,364
gqa_formal	train	330,194
gqa_informal	train	284,650
qa_balance	train	300,875
qa_informal	train	411,033

Provenance

The corpus derives from 2,025,730 source rows expanded by language and style. Gemma 3 27B IT produced Persian→English/Arabic and English→Chinese/Hebrew/Spanish variants. The stage used all six answer-translation groups plus complete, moderate, formal and informal answering exports. The 329 construction components combine into 22 final files. 1,961 complete-answer records have known language-directive conflicts; complete QA covers 27 cross-language directions.

This release is part of HexaBee-4B, previously labeled PCT-4B in the local evaluation audit. Both names refer to the same final translator. Renaming does not imply retraining. This repository is an artifact release, not evidence of workshop acceptance or an official shared-task ranking.

Limitations

Historical exports may contain empty strings, imperfect generated text and instruction conflicts. Source row reuse is not independent evidence. See measured release statistics and the accompanying evaluation scope.

The complete adaptation recipe was evaluated; no matched ablation isolates the contribution of multilingual answering. Synthetic references and model-based judges limit interpretation. Long COMET inputs are truncated at the encoder limit. Use the independent evaluation repository for precise sample counts, parameters, all directions, and uncertainty estimates.

Release family

Artifact	Repository
HexaBee-4B	Iman998/HexaBee-4B
HexaBee · Multilingual Base	Iman998/HexaBee-4B-Multilingual-Base
HexaBee · Wikipedia PT LoRA	Iman998/HexaBee-4B-PT-LoRA
HexaBee · Multilingual SFT LoRA	Iman998/HexaBee-4B-Multilingual-SFT-LoRA
HexaBee · Translation LoRA	Iman998/HexaBee-4B-Translation-LoRA
HexaBee · Wikipedia Foundations	Iman998/HexaBee-Wikipedia-PT
HexaBee · Multilingual Answering & Translation	Iman998/HexaBee-Multilingual-SFT
HexaBee · Bidirectional Translation	Iman998/HexaBee-Bidirectional-Translation
HexaBee · The Prompt Hive	Iman998/HexaBee-System-Prompts
HexaBee · Translation Archive	Iman998/HexaBee-Model-Outputs
HexaBee · Evaluation Atlas	Iman998/HexaBee-Evaluation

Each phase, checkpoint, prompt inventory, output archive and evaluation result has its own repository and card. Model cards identify the exact adapter base; dataset cards identify the unit of a released row. Historical row counts are not counts of unique source articles.

License and attribution

See LICENSE.md. Model weights remain under Gemma terms; upstream dataset and service terms are not replaced by this repository's software license. The illustrated branding is decorative and does not encode experimental evidence.

Citation and references

Please cite Barati, Rostamkhani, and Badzohreh (2026), HexaBee-4B: Six-Language Translation with Multilingual Answering and Bidirectional Supervision. Research code.

Related methods and tools: Gemma 3, LoRA, LLaMA-Factory, SacreBLEU, COMET.

---

⬡ HexaBee-4B · FA · EN · AR · ZH · HE · ES
Maintained by Iman998

Cite this paper

@misc{barati2026hexabee4b,
  title={HexaBee-4B: Six-Language Translation with Multilingual Answering and Bidirectional Supervision},
  author={Barati, Iman and Rostamkhani, Mohammadmostafa and Badzohreh, Mohammad},
  year={2026},
  eprint={2610.16361},
  archivePrefix={arXiv},
  url={https://arxiv.org/abs/2610.16361}
}