HexaBee-4B: Six-Language Translation with Multilingual Answering and Bidirectional Supervision

• Translation must preserve source meaning while producing text in a specified target language.Source: §1 · p. 1
• Multilingual instruction data can support cross-language generalization, but separating question and answer languages makes cross-language answering a distinct task.Source: §1 · p. 1
• A multilingual answer can supervise both a question asked in one language and translation of an answer text into the requested language.Source: §1 · p. 1
• HEXABEE-4B combines Wikipedia language modeling, multilingual response supervision, and translation specialization, then evaluates the complete recipe against public Gemma 3 4B IT and 12B IT.The comparison does not isolate the effect of an individual training component.Source: §1 · p. 2

• Multilingual MT uses shared parameters and explicit target-language signals (Johnson et al., 2017), while prior work studies continued training, translation instructions, and teacher-generated translations.Source: §2 · p. 2
• Multilingual instruction resources and cross-language instruction–response pairs provide precedents for coupling instruction following with translation (Muennighoff et al., 2023; Lin et al., 2025).Source: §2 · p. 2
• Unlike the cited QA benchmarks, this paper uses explicit language roles as generation supervision for a translator and evaluates translation quality rather than answering or knowledge composition.Source: §2 · p. 2

3.1 Answering and translation as connected tasks

• The adaptation recipe first trains on Wikipedia, continues the same adapter on multilingual answering and answer-text translation, then trains a fresh adapter on bidirectional translations.Source: Figure 1 · p. 3
• A shared answer can serve as the target for both QA, whose input is a question, and translation, whose input is answer text.Source: §3.1 · p. 3
• Instruction pools vary response language, extent, style, and translation direction; the wording language of a directive is separate from its requested response language.Source: §3.1 · p. 3

3.2 A shared response objective

• QA and translation examples use the same response-token negative log-likelihood, with instructions and inputs as conditioning text and the response as supervision.Source: §3.2 · p. 3
• The phase-2 dataset is the multiset union of retained QA and translation records, without a separate alignment loss between the tasks.Source: §3.2 · pp. 3–4

4.1 Phase 1: Wikipedia continued pretraining

• Phase 1 continues pretraining on monolingual Wikipedia text in English, Persian, Arabic, Chinese, and Egyptian Arabic before paired generation tasks begin.A one-epoch LoRA adapter trained on 3,399,664 processed training examples is retained for phase 2.Source: §4.1 · p. 4

4.2 Phase 2: multilingual answering and answer translation

• Phase 2 combines question-to-answer supervision with translation between answer variants associated with the same content.Its retained mixture contains 19,458,332 answering records and 10,695,072 answer-text translations; 5,759,806 answering records cross question and response languages.Source: §4.2 · p. 4
• The phase-2 QA distribution is asymmetric: complete-answer data cover 27 cross-language directions, while combined translation data cover all 30.The retained corpus also includes empty fields and a small set of conflicting language directives.Source: §4.2 · p. 4

4.3 Phase 3: bidirectional translation specialization

• Phase 3 uses a separate teacher-generated corpus spanning topics, genres, registers, and complexity levels, but its synthetic test split measures that particular distribution.The authors state that diversity is a construction goal, not evidence of coverage of every deployment domain.Source: §4.3 · p. 5
• The final phase trains on 658,274 bidirectional examples from 329,137 retained synthetic pairs, with each pair used in both translation directions.The reverse example reuses the existing texts rather than requiring another teacher call.Source: §4.3 · p. 5

5.1 Evaluation data and comparison systems

• The comparison uses 36,572 held-out synthetic examples across 30 directions, with all systems scored against the same source texts and references; 300 examples receive model judgments.The judged subset contains ten examples per direction.Source: §5.1 · pp. 5–6
• The archived inference templates differ, so the comparison reflects the recorded systems and prompts rather than adaptation under identical prompts; no human-authored benchmark is evaluated.Outputs are scored as stored, including empty responses and unsolicited text.Source: §5.1 · p. 6

5.2 Automatic metrics

• Automatic evaluation reports BLEU, chrF, chrF++, TER, and COMET, with equal weight assigned to each translation direction in the primary summary.TER is lower-is-better; the other metrics are higher-is-better, and COMET values are multiplied by 100 for display.Source: §5.2 · p. 6

5.3 Model judging and uncertainty

• The archived DeepSeek judge scores translations reference-assisted or reference-free, producing holistic scores and eight rubric dimensions; these are model judgments, not calibrated accuracy percentages or human annotations.Paired bootstrap intervals on the 300-example subset capture item-sampling uncertainty, not variability across training runs or alternative judges.Source: §5.3 · p. 6

6.2 Language and direction differences

• chrF++ improves in every evaluated direction against both baselines, with gains extending across target languages.Figure 4 shows improvements in both direction-macro chrF++ and COMET for each target language.Source: §6.2 · p. 6, Figure 4 · p. 7
• 90.09 TER for Chinese-to-Hebrew is worse than 88.40 for Gemma 3 4B IT and 78.13 for Gemma 3 12B IT, despite higher BLEU and chrF++.TER improves in the other 29 directions. Lower absolute Chinese-target chrF++ is not a calibrated difficulty ranking because scripts, segmentation, references, and metric components affect cross-language comparisons.Source: §6.2 · p. 6

6.1 Overall translation quality

• 73.75 direction-macro chrF++ gives HEXABEE-4B the top overall score across five automatic metrics, versus 57.11 for Gemma 3 4B IT and 66.83 for Gemma 3 12B IT.The comparison covers the full test set; Table 2 reports direction-macro automatic metrics.Source: §6.1 · p. 6, Table 2 · p. 7

6.3 Judged quality

• Both judge modes follow the automatic-metric ranking on the matched subset, with a considerably closer comparison to 12B IT than to 4B IT.Table 3 reports paired differences; against 4B IT, HEXABEE-4B wins, ties, or loses on 214, 54, and 32 reference-assisted examples. Terminology scores 90.42 and remains the lowest reference-assisted dimension.Source: Table 3 · p. 7, §6.3 · p. 7

7 Discussion

• Continued pretraining adapts language modeling to encyclopedic text, while multilingual answering and answer-text translation provide intermediate supervision before bidirectional translation specialization.Answering selects requested information and expresses it in a specified language; answer-text translation exposes semantic counterparts in that response space.Source: §7 · p. 7
• Shared answer targets connect intermediate answering and translation, while the final stage uses a separate source-generation process.Source: §7 · p. 7
• The complete recipe is reported to produce a stronger translator on the recorded test set, but the results do not attribute gains to individual components.Source: §7 · pp. 7–8
• A controlled study would fix the adapted checkpoint, training budget, final corpus, and prompts while varying the intermediate mixture; those comparisons are unavailable.Source: §7 · p. 8

8 Conclusion

• HEXABEE-4B adapts Gemma 3 4B IT through Wikipedia continued pretraining, multilingual answering with answer-text translation, and bidirectional translation specialization.The intermediate stage links tasks through multilingual answer targets, including 5.76 million cross-language QA records in a 30.15-million-record mixture.Source: §8 · p. 8
• On 36,572 synthetic test examples, HEXABEE-4B improves all reported aggregate metrics over Gemma 3 4B IT and 12B IT, with chrF++ gains in every direction.The test spans 30 directions among Arabic, Chinese, English, Hebrew, Persian, and Spanish.Source: §8 · p. 8
• The study presents a reproducible translation-focused adaptation recipe and a direction-level account of its strengths and remaining weaknesses.The conclusion describes these as the study's supported contribution and account.Source: §8 · p. 8

Limitations

• Synthetic test sources and references, plus a judge from the same teacher-service family, leave human-authored and out-of-domain quality unestablished.The passage says teacher-like outputs may be favored.Source: Limitations · p. 9
• Record-level splitting permits topic overlap, and the audit does not establish the absence of semantic duplicates or pretraining exposure.Two source strings overlap after whitespace normalization; the audit does not establish semantic deduplication.Source: Limitations · p. 9
• Without human assessment, controlled mixture ablation, or intermediate-checkpoint evaluation, the study cannot isolate causes of the gains.Prompt differences across systems and context-limit failures also constrain interpretation of the comparison.Source: Limitations · p. 9

Ethical Considerations

• Generated data may unevenly represent languages and topics, alter names or facts, and have unvalidated high-stakes translation quality; licensing obligations and incomplete article provenance also constrain use.The paper reports existing artifacts without new human annotations and does not claim blanket licensing over all text.Source: Ethical Considerations · p. 9

A Continued-pretraining corpus

• 3,399,664 phase-1 examples are reported after preprocessing; raw source-row totals are neither unique-article counts nor processed-sequence counts.Four Persian registrations refer to one source file, counted once in the raw-source table.Source: A Continued-pretraining corpus · p. 11, Table 4 · p. 11

B.1 Multilingual answer inventory

• 10,695,072 answer-text translations remain after null filtering, while the released phase-2 mixture contains 30,153,404 training records across tasks.The inventory includes 5,759,806 cross-language QA records; generated answer variants have not all been exhaustively human-checked.Source: §B.1 · p. 11

B.2 System-instruction composition

• Answer-language directives are distinct from the language used to phrase them; completeness, answer-language, and optional style directives are sampled and combined.Duplicates remain in prompt pools, so repeated entries retain sampling mass; seed 42 governs construction choices.Source: §B.2 · pp. 11–12

B.3 Synthetic translation generation

• 365,718 synthetic translation pairs are split into 329,146 raw training and 36,572 test pairs, with no split seed recorded.Filtering yields 329,137 forward and 329,137 reverse training records; the teacher instructions do not guarantee translation quality.Source: §B.3 · p. 12

B.4 Archived inference

• All systems use temperature-zero decoding, but their prompts differ; the judge uses reference-assisted and reference-free scoring, with requested caps rather than verified error labels.The judge subset has ten examples per direction; some baseline empty responses follow context-limit failures.Source: §B.4 · p. 12, §B.5 · p. 12

C Phase-2 corpus composition

• The phase-2 export retains empty fields and includes 1,961 Spanish-target complete-answer records whose directives incorrectly request English.The conflicts cover all cross-language Spanish-target complete-answer records, so correctly specified Spanish cross-language answering is not established by this subset.Source: C Phase-2 corpus composition · pp. 12–13, Table 5 · p. 13

D Prompt sampling details

• Prompt sampling normalizes configured weights over available language pools and samples rows uniformly; final-stage coverage must be measured from instantiated prompts.The audit finds zero exact raw-string train–test overlaps but two whitespace-normalized matches, and does not establish absence of semantic duplicates or pretraining exposure.Source: D Prompt sampling details · p. 13, §D.1 · p. 14, §D.2 · p. 15

E Training and serialization

• Text-only examples serialize a system instruction and source text in the user turn, followed by the target text in the model turn.The training configuration uses 4,096-token sequences and checkpoints every 600 steps.Source: E Training and serialization · p. 15
• Final-stage adaptation affects text projections only, while the multimodal checkpoint retains its vision components.The nominal 4B size identifies the Gemma model family.Source: E Training and serialization · p. 15
• The translation stage starts a new adapter on the merged cross-lingual model; phase 2 combines system-directed answering with multilingual answer-text translations, and phase 3 specializes the merged model on independent synthetic translations.The saved adapters use rank 256, scale 512, and dropout 0.005.Source: E Training and serialization · p. 15, Figure 7 · p. 16

F Metrics and results by language

• Target-language summaries give the five incoming directions equal weight, whereas direction-macro COMET averages the 30 direction means.Source: F Metrics and results by language · pp. 16–17
• Pooled-corpus chrF++ is 77.30 for HEXABEE-4B, 71.13 for Gemma 3 12B IT, and 60.77 for Gemma 3 4B IT.These values differ from direction-macro scores.Source: F Metrics and results by language · p. 16
• chrF++ uses character orders 1–6 and word orders 1–2, while chrF uses only character orders; corpus counts are aggregated before the harmonic combination.Source: F Metrics and results by language · p. 16
• TER uses Tercom-style edit search with shifts, an approximation rather than a certified global minimum; corpus and macro aggregation remain distinct.The equation counts substitutions, deletions, insertions, and shifts against reference-token count.Source: F Metrics and results by language · p. 17
• Judge summaries cover eight dimensions on matched items (300 overall, 50 per target), but remain judgments from one evaluator, not independent human assessments.Source: Figure 9 · p. 18, Table 11 · p. 19, Table 24 · p. 24

G Evaluation integrity and reproducibility

• Reparsing original judge JSON recovered 11 reference-view and 13 reference-free zeros for the 4B baseline, plus one reference-free zero for the 12B baseline; all 300 items then had valid scores in both views.The reparsed means agree with the final flattened score files.Source: G Evaluation integrity and… · p. 25
• Content hashes identify paired outputs because row indices are unstable after asynchronous processing and sampling changes order; an explicit JSON score of 0 remains zero.Source: G Evaluation integrity and… · p. 25
• The paper retrospectively analyzes existing translation outputs and newly computed metrics, rather than reporting new training runs or newly collected human labels.Source: G Evaluation integrity and… · p. 25