Iman998/HexaBee-Evaluation

HexaBee · Evaluation Atlas

Thirty directions. Three systems. Traceable evidence.

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

All translation directions, automatic metrics, eight-dimension model judging, paired bootstrap intervals, and the data-composition audit in one independently downloadable evaluation repository.

Property	Value
Artifact	HexaBee · Evaluation Atlas
Scope	Numeric results · three systems · 30 directions
Tabular results	90 direction rows + 900 paired judge rows
Format	CSV / JSON + figures
Release	Audited historical artifacts
Model family	HexaBee-4B / Gemma 3

Translation benchmark

System	BLEU ↑	chrF ↑	chrF++ ↑	TER ↓	COMET ×100 ↑
Gemma 3 4B IT	40.70	58.91	57.11	51.40	85.88
Gemma 3 12B IT	50.54	68.70	66.83	38.89	88.20
HexaBee-4B	62.03	75.08	73.75	30.05	89.86

Equal-weight macro over 30 directions, 36,572 test items per model. Best displayed values are bold; the second distinct value is underlined. TER is lower-is-better. Synthetic references and differing archived inference templates limit comparisons. These are archived research results, not an official shared-task ranking.

Quick start

from datasets import load_dataset

ds = load_dataset("Iman998/HexaBee-Evaluation", "direction_metrics", split="test", streaming=True)
print(next(iter(ds)))

Contents

all_direction_metrics.csv: 90 system/direction rows with BLEU, chrF, chrF++, TER, COMET, judge means and all eight rubric dimensions in both modes. paired_scores.csv: 900 score-only rows for paired intervals. Additional JSON files contain metric signatures, sufficient statistics, COMET checkpoint provenance, and audited phase-2 data counts. Figures are included in PNG, SVG and PDF formats. The expandable tables below compare all models separately for each target language.

release_manifest.json lists distributed files, sizes, and SHA-256. release_stats.json records measured artifact metadata.

Available configurations

Configuration	Splits	Rows
direction_metrics	test	90
paired_judge_scores	test	900

Results by target language

Automatic metrics below use the full test. The separate judge tables use ten matched items per direction and mode.

Persian · five incoming directions

Target macro

System	BLEU ↑	chrF ↑	chrF++ ↑	TER ↓	COMET ×100 ↑
Gemma 3 4B IT	32.44	57.57	55.67	59.16	84.81
Gemma 3 12B IT	43.61	70.24	68.22	41.59	88.32
HexaBee-4B	57.43	77.34	75.90	31.68	89.91

Arabic-Persian · n = 1,228 per model

System	BLEU ↑	chrF ↑	chrF++ ↑	TER ↓	COMET ×100 ↑
Gemma 3 4B IT	32.96	57.46	55.29	56.93	84.60
Gemma 3 12B IT	45.58	71.82	69.85	40.84	88.30
HexaBee-4B	62.29	80.41	79.05	23.95	90.46

Judge subset · n = 10 per model, each mode

System	Reference-aware ↑	Reference-free ↑
Gemma 3 4B IT	73.20	75.10
Gemma 3 12B IT	89.80	92.60
HexaBee-4B	89.50	95.10

Chinese-Persian · n = 1,208 per model

System	BLEU ↑	chrF ↑	chrF++ ↑	TER ↓	COMET ×100 ↑
Gemma 3 4B IT	19.17	41.95	40.29	90.09	81.51
Gemma 3 12B IT	32.94	63.68	61.27	55.33	87.85
HexaBee-4B	45.05	69.11	67.38	55.17	89.04

Judge subset · n = 10 per model, each mode

System	Reference-aware ↑	Reference-free ↑
Gemma 3 4B IT	51.60	52.70
Gemma 3 12B IT	89.40	91.50
HexaBee-4B	91.00	93.10

English-Persian · n = 1,239 per model

System	BLEU ↑	chrF ↑	chrF++ ↑	TER ↓	COMET ×100 ↑
Gemma 3 4B IT	34.35	62.41	60.48	52.63	86.10
Gemma 3 12B IT	42.64	68.48	66.60	41.52	88.24
HexaBee-4B	56.25	76.24	74.81	29.15	90.13

Judge subset · n = 10 per model, each mode

System	Reference-aware ↑	Reference-free ↑
Gemma 3 4B IT	89.70	88.80
Gemma 3 12B IT	91.70	91.90
HexaBee-4B	93.00	93.80

Hebrew-Persian · n = 1,198 per model

System	BLEU ↑	chrF ↑	chrF++ ↑	TER ↓	COMET ×100 ↑
Gemma 3 4B IT	40.23	64.44	62.67	44.97	86.19
Gemma 3 12B IT	50.15	74.71	72.89	33.55	88.98
HexaBee-4B	63.73	81.50	80.26	23.03	90.18

Judge subset · n = 10 per model, each mode

System	Reference-aware ↑	Reference-free ↑
Gemma 3 4B IT	90.40	91.80
Gemma 3 12B IT	93.40	94.80
HexaBee-4B	94.50	95.90

Spanish-Persian · n = 1,196 per model

System	BLEU ↑	chrF ↑	chrF++ ↑	TER ↓	COMET ×100 ↑
Gemma 3 4B IT	35.51	61.60	59.63	51.17	85.67
Gemma 3 12B IT	46.74	72.49	70.47	36.71	88.22
HexaBee-4B	59.82	79.44	78.03	27.11	89.71

Judge subset · n = 10 per model, each mode

System	Reference-aware ↑	Reference-free ↑
Gemma 3 4B IT	79.40	80.90
Gemma 3 12B IT	90.90	92.20
HexaBee-4B	92.30	93.80

English · five incoming directions

Target macro

System	BLEU ↑	chrF ↑	chrF++ ↑	TER ↓	COMET ×100 ↑
Gemma 3 4B IT	55.28	74.13	72.40	36.35	87.40
Gemma 3 12B IT	67.16	84.77	83.36	22.59	88.90
HexaBee-4B	75.06	87.68	86.76	17.00	89.83

Arabic-English · n = 1,200 per model

System	BLEU ↑	chrF ↑	chrF++ ↑	TER ↓	COMET ×100 ↑
Gemma 3 4B IT	59.91	75.60	74.32	29.38	86.98
Gemma 3 12B IT	70.24	86.13	84.97	21.03	88.74
HexaBee-4B	76.36	88.32	87.58	16.02	89.58

Judge subset · n = 10 per model, each mode

System	Reference-aware ↑	Reference-free ↑
Gemma 3 4B IT	82.00	83.30
Gemma 3 12B IT	92.40	94.80
HexaBee-4B	90.10	94.60

Chinese-English · n = 1,213 per model

System	BLEU ↑	chrF ↑	chrF++ ↑	TER ↓	COMET ×100 ↑
Gemma 3 4B IT	38.89	64.46	62.03	59.53	86.20
Gemma 3 12B IT	47.49	75.48	72.87	40.65	87.93
HexaBee-4B	58.71	79.80	77.87	32.67	89.16

Judge subset · n = 10 per model, each mode

System	Reference-aware ↑	Reference-free ↑
Gemma 3 4B IT	64.40	65.90
Gemma 3 12B IT	95.70	94.90
HexaBee-4B	95.20	95.60

Hebrew-English · n = 1,185 per model

System	BLEU ↑	chrF ↑	chrF++ ↑	TER ↓	COMET ×100 ↑
Gemma 3 4B IT	63.49	77.80	76.72	28.99	87.63
Gemma 3 12B IT	73.72	87.81	86.82	16.72	89.22
HexaBee-4B	79.87	90.07	89.47	12.54	90.03

Judge subset · n = 10 per model, each mode

System	Reference-aware ↑	Reference-free ↑
Gemma 3 4B IT	93.40	92.40
Gemma 3 12B IT	91.80	89.20
HexaBee-4B	90.80	90.00

Persian-English · n = 1,234 per model

System	BLEU ↑	chrF ↑	chrF++ ↑	TER ↓	COMET ×100 ↑
Gemma 3 4B IT	57.78	76.96	75.53	29.13	88.46
Gemma 3 12B IT	67.46	84.48	83.21	20.09	89.29
HexaBee-4B	76.31	88.01	87.22	14.41	90.16

Judge subset · n = 10 per model, each mode

System	Reference-aware ↑	Reference-free ↑
Gemma 3 4B IT	95.10	95.40
Gemma 3 12B IT	96.60	95.80
HexaBee-4B	96.70	96.70

Spanish-English · n = 1,178 per model

System	BLEU ↑	chrF ↑	chrF++ ↑	TER ↓	COMET ×100 ↑
Gemma 3 4B IT	56.34	75.84	73.42	34.73	87.71
Gemma 3 12B IT	76.88	89.96	88.91	14.43	89.31
HexaBee-4B	84.06	92.21	91.63	9.37	90.24

Judge subset · n = 10 per model, each mode

System	Reference-aware ↑	Reference-free ↑
Gemma 3 4B IT	96.40	97.20
Gemma 3 12B IT	97.40	97.40
HexaBee-4B	94.80	96.40

Arabic · five incoming directions

Target macro

System	BLEU ↑	chrF ↑	chrF++ ↑	TER ↓	COMET ×100 ↑
Gemma 3 4B IT	34.09	62.25	59.33	56.99	86.64
Gemma 3 12B IT	43.14	70.43	67.69	44.60	88.52
HexaBee-4B	56.17	77.09	75.10	33.62	90.17

Chinese-Arabic · n = 1,229 per model

System	BLEU ↑	chrF ↑	chrF++ ↑	TER ↓	COMET ×100 ↑
Gemma 3 4B IT	23.68	51.49	48.28	80.63	84.07
Gemma 3 12B IT	32.53	64.44	60.91	61.20	87.91
HexaBee-4B	42.46	69.06	66.34	57.30	89.15

Judge subset · n = 10 per model, each mode

System	Reference-aware ↑	Reference-free ↑
Gemma 3 4B IT	72.60	71.60
Gemma 3 12B IT	91.60	92.50
HexaBee-4B	90.40	93.00

English-Arabic · n = 1,213 per model

System	BLEU ↑	chrF ↑	chrF++ ↑	TER ↓	COMET ×100 ↑
Gemma 3 4B IT	37.50	66.77	63.77	51.70	87.99
Gemma 3 12B IT	42.30	68.56	66.09	47.70	87.47
HexaBee-4B	55.60	76.61	74.66	35.72	90.06

Judge subset · n = 10 per model, each mode

System	Reference-aware ↑	Reference-free ↑
Gemma 3 4B IT	86.50	85.20
Gemma 3 12B IT	81.10	82.20
HexaBee-4B	93.40	94.10

Hebrew-Arabic · n = 1,230 per model

System	BLEU ↑	chrF ↑	chrF++ ↑	TER ↓	COMET ×100 ↑
Gemma 3 4B IT	39.08	63.41	61.23	48.02	86.42
Gemma 3 12B IT	51.01	74.60	72.25	32.19	89.30
HexaBee-4B	63.65	80.99	79.44	23.26	90.59

Judge subset · n = 10 per model, each mode

System	Reference-aware ↑	Reference-free ↑
Gemma 3 4B IT	78.30	74.70
Gemma 3 12B IT	88.40	87.90
HexaBee-4B	92.60	93.60

Persian-Arabic · n = 1,228 per model

System	BLEU ↑	chrF ↑	chrF++ ↑	TER ↓	COMET ×100 ↑
Gemma 3 4B IT	32.88	63.15	60.05	56.62	87.51
Gemma 3 12B IT	46.62	72.69	70.14	38.53	89.75
HexaBee-4B	61.00	79.59	78.00	24.43	91.27

Judge subset · n = 10 per model, each mode

System	Reference-aware ↑	Reference-free ↑
Gemma 3 4B IT	88.70	89.80
Gemma 3 12B IT	92.60	93.40
HexaBee-4B	94.40	95.10

Spanish-Arabic · n = 1,220 per model

System	BLEU ↑	chrF ↑	chrF++ ↑	TER ↓	COMET ×100 ↑
Gemma 3 4B IT	37.28	66.41	63.33	47.96	87.21
Gemma 3 12B IT	43.22	71.87	69.06	43.37	88.15
HexaBee-4B	58.14	79.19	77.07	27.40	89.78

Judge subset · n = 10 per model, each mode

System	Reference-aware ↑	Reference-free ↑
Gemma 3 4B IT	86.70	86.70
Gemma 3 12B IT	91.90	92.90
HexaBee-4B	92.80	92.20

Chinese · five incoming directions

Target macro

System	BLEU ↑	chrF ↑	chrF++ ↑	TER ↓	COMET ×100 ↑
Gemma 3 4B IT	35.22	31.87	30.97	60.19	84.52
Gemma 3 12B IT	40.15	38.40	36.57	54.28	87.01
HexaBee-4B	54.78	50.31	48.67	38.68	90.00

Arabic-Chinese · n = 1,201 per model

System	BLEU ↑	chrF ↑	chrF++ ↑	TER ↓	COMET ×100 ↑
Gemma 3 4B IT	36.56	32.17	30.93	58.04	83.81
Gemma 3 12B IT	42.36	39.84	37.13	50.06	86.94
HexaBee-4B	57.09	51.82	49.12	35.01	89.69

Judge subset · n = 10 per model, each mode

System	Reference-aware ↑	Reference-free ↑
Gemma 3 4B IT	73.60	73.90
Gemma 3 12B IT	91.80	91.90
HexaBee-4B	93.70	94.20

English-Chinese · n = 1,276 per model

System	BLEU ↑	chrF ↑	chrF++ ↑	TER ↓	COMET ×100 ↑
Gemma 3 4B IT	30.42	27.90	27.56	69.42	83.44
Gemma 3 12B IT	32.77	31.48	29.94	67.91	85.07
HexaBee-4B	50.33	45.82	45.30	45.39	90.02

Judge subset · n = 10 per model, each mode

System	Reference-aware ↑	Reference-free ↑
Gemma 3 4B IT	82.30	82.80
Gemma 3 12B IT	91.80	93.50
HexaBee-4B	93.00	93.60

Hebrew-Chinese · n = 1,202 per model

System	BLEU ↑	chrF ↑	chrF++ ↑	TER ↓	COMET ×100 ↑
Gemma 3 4B IT	41.02	38.25	37.22	51.04	86.17
Gemma 3 12B IT	44.57	42.44	40.97	45.95	87.67
HexaBee-4B	58.14	53.41	51.80	33.49	89.75

Judge subset · n = 10 per model, each mode

System	Reference-aware ↑	Reference-free ↑
Gemma 3 4B IT	76.10	77.20
Gemma 3 12B IT	93.00	93.50
HexaBee-4B	94.20	94.30

Persian-Chinese · n = 1,251 per model

System	BLEU ↑	chrF ↑	chrF++ ↑	TER ↓	COMET ×100 ↑
Gemma 3 4B IT	38.95	35.74	34.01	53.12	86.19
Gemma 3 12B IT	45.23	42.86	40.58	44.42	88.67
HexaBee-4B	57.23	52.86	50.30	34.91	90.41

Judge subset · n = 10 per model, each mode

System	Reference-aware ↑	Reference-free ↑
Gemma 3 4B IT	74.70	75.00
Gemma 3 12B IT	92.10	90.50
HexaBee-4B	94.40	94.60

Spanish-Chinese · n = 1,205 per model

System	BLEU ↑	chrF ↑	chrF++ ↑	TER ↓	COMET ×100 ↑
Gemma 3 4B IT	29.14	25.27	25.15	69.33	83.01
Gemma 3 12B IT	35.84	35.37	34.24	63.07	86.73
HexaBee-4B	51.11	47.63	46.84	44.61	90.14

Judge subset · n = 10 per model, each mode

System	Reference-aware ↑	Reference-free ↑
Gemma 3 4B IT	69.40	71.10
Gemma 3 12B IT	80.90	84.70
HexaBee-4B	83.80	86.30

Hebrew · five incoming directions

Target macro

System	BLEU ↑	chrF ↑	chrF++ ↑	TER ↓	COMET ×100 ↑
Gemma 3 4B IT	31.94	54.98	53.02	59.27	85.87
Gemma 3 12B IT	43.32	64.53	62.78	46.14	88.17
HexaBee-4B	54.71	70.90	69.90	42.17	89.68

Arabic-Hebrew · n = 1,255 per model

System	BLEU ↑	chrF ↑	chrF++ ↑	TER ↓	COMET ×100 ↑
Gemma 3 4B IT	33.86	55.36	53.52	55.20	85.90
Gemma 3 12B IT	48.39	68.91	67.01	37.34	88.91
HexaBee-4B	61.41	76.53	75.50	28.87	91.03

Judge subset · n = 10 per model, each mode

System	Reference-aware ↑	Reference-free ↑
Gemma 3 4B IT	70.40	66.30
Gemma 3 12B IT	91.70	93.20
HexaBee-4B	93.00	94.70

Chinese-Hebrew · n = 1,232 per model

System	BLEU ↑	chrF ↑	chrF++ ↑	TER ↓	COMET ×100 ↑
Gemma 3 4B IT	19.81	39.55	38.42	88.40	82.50
Gemma 3 12B IT	27.88	48.10	47.27	78.13	85.34
HexaBee-4B	35.82	51.34	51.36	90.09	86.74

Judge subset · n = 10 per model, each mode

System	Reference-aware ↑	Reference-free ↑
Gemma 3 4B IT	85.00	87.00
Gemma 3 12B IT	90.30	88.60
HexaBee-4B	93.70	94.70

English-Hebrew · n = 1,216 per model

System	BLEU ↑	chrF ↑	chrF++ ↑	TER ↓	COMET ×100 ↑
Gemma 3 4B IT	36.21	61.46	59.13	49.77	87.17
Gemma 3 12B IT	44.79	65.92	64.17	41.62	88.31
HexaBee-4B	56.73	73.34	72.04	32.13	89.85

Judge subset · n = 10 per model, each mode

System	Reference-aware ↑	Reference-free ↑
Gemma 3 4B IT	71.20	77.20
Gemma 3 12B IT	92.60	92.10
HexaBee-4B	93.40	92.80

Persian-Hebrew · n = 1,238 per model

System	BLEU ↑	chrF ↑	chrF++ ↑	TER ↓	COMET ×100 ↑
Gemma 3 4B IT	33.83	58.93	56.54	52.35	87.53
Gemma 3 12B IT	45.84	69.08	66.79	38.09	89.74
HexaBee-4B	58.38	76.09	74.67	33.36	90.94

Judge subset · n = 10 per model, each mode

System	Reference-aware ↑	Reference-free ↑
Gemma 3 4B IT	70.40	75.60
Gemma 3 12B IT	90.10	91.20
HexaBee-4B	92.10	94.50

Spanish-Hebrew · n = 1,255 per model

System	BLEU ↑	chrF ↑	chrF++ ↑	TER ↓	COMET ×100 ↑
Gemma 3 4B IT	35.99	59.59	57.50	50.63	86.26
Gemma 3 12B IT	49.71	70.65	68.68	35.51	88.55
HexaBee-4B	61.20	77.18	75.92	26.40	89.85

Judge subset · n = 10 per model, each mode

System	Reference-aware ↑	Reference-free ↑
Gemma 3 4B IT	83.00	86.10
Gemma 3 12B IT	90.70	91.90
HexaBee-4B	93.60	94.80

Spanish · five incoming directions

Target macro

System	BLEU ↑	chrF ↑	chrF++ ↑	TER ↓	COMET ×100 ↑
Gemma 3 4B IT	55.21	72.69	71.24	36.44	86.02
Gemma 3 12B IT	65.85	83.83	82.35	24.13	88.29
HexaBee-4B	74.04	87.18	86.16	17.13	89.57

Arabic-Spanish · n = 1,257 per model

System	BLEU ↑	chrF ↑	chrF++ ↑	TER ↓	COMET ×100 ↑
Gemma 3 4B IT	59.24	74.94	73.59	30.24	85.60
Gemma 3 12B IT	68.19	85.36	84.08	20.85	87.93
HexaBee-4B	77.23	88.55	87.73	14.41	89.37

Judge subset · n = 10 per model, each mode

System	Reference-aware ↑	Reference-free ↑
Gemma 3 4B IT	88.20	91.10
Gemma 3 12B IT	94.10	96.30
HexaBee-4B	91.80	94.20

Chinese-Spanish · n = 1,173 per model

System	BLEU ↑	chrF ↑	chrF++ ↑	TER ↓	COMET ×100 ↑
Gemma 3 4B IT	32.40	55.60	53.72	68.49	83.08
Gemma 3 12B IT	48.63	75.81	73.37	44.50	87.27
HexaBee-4B	58.16	80.03	78.11	31.49	89.03

Judge subset · n = 10 per model, each mode

System	Reference-aware ↑	Reference-free ↑
Gemma 3 4B IT	81.90	86.20
Gemma 3 12B IT	93.60	92.90
HexaBee-4B	94.10	94.90

English-Spanish · n = 1,188 per model

System	BLEU ↑	chrF ↑	chrF++ ↑	TER ↓	COMET ×100 ↑
Gemma 3 4B IT	66.60	84.25	82.63	23.34	88.08
Gemma 3 12B IT	73.09	87.64	86.37	17.24	88.72
HexaBee-4B	79.33	90.12	89.27	12.45	89.61

Judge subset · n = 10 per model, each mode

System	Reference-aware ↑	Reference-free ↑
Gemma 3 4B IT	93.20	94.90
Gemma 3 12B IT	94.30	95.90
HexaBee-4B	93.90	96.20

Hebrew-Spanish · n = 1,183 per model

System	BLEU ↑	chrF ↑	chrF++ ↑	TER ↓	COMET ×100 ↑
Gemma 3 4B IT	58.27	71.77	70.79	30.88	85.57
Gemma 3 12B IT	71.49	85.66	84.60	17.18	88.34
HexaBee-4B	79.39	89.05	88.43	11.91	89.55

Judge subset · n = 10 per model, each mode

System	Reference-aware ↑	Reference-free ↑
Gemma 3 4B IT	78.40	77.20
Gemma 3 12B IT	96.00	86.70
HexaBee-4B	96.50	97.00

Persian-Spanish · n = 1,241 per model

System	BLEU ↑	chrF ↑	chrF++ ↑	TER ↓	COMET ×100 ↑
Gemma 3 4B IT	59.54	76.89	75.45	29.25	87.78
Gemma 3 12B IT	67.83	84.70	83.34	20.90	89.22
HexaBee-4B	76.09	88.16	87.27	15.38	90.28

Judge subset · n = 10 per model, each mode

System	Reference-aware ↑	Reference-free ↑
Gemma 3 4B IT	83.50	89.90
Gemma 3 12B IT	94.20	92.30
HexaBee-4B	94.90	93.30

All eight judge dimensions in both modes are retained in the direction-level CSV. Paired intervals use 10,000 direction-stratified bootstrap replicates.

Provenance

Full test: 36,572 examples per model. Judge subset: 300 matched examples per model, ten per direction. SacreBLEU 2.5.1; BLEU zh tokenizer for Chinese targets and 13a otherwise; chrF character order 6; chrF++ word order 2; TER normalized with Asian support; unbabel-comet 2.2.7 and wmt22-comet-da (512-position encoder). Judge intervals use 10,000 paired direction-stratified bootstrap replicates and seed 20260912. Zero scores were recovered from raw responses rather than discarded.

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}
}