Language modelOpen weights
Index-Translate-2B
Bilibili Index Team
- Released
- September 30, 2026
- Data date
- October 7, 2026
- Input
- Text and translation instructions
- Output
- Translated text
Index-Translate-2B is built for text translation with explicit constraints on terminology, style, and output format. Its model card lists 150 languages and also presents the translation of community expressions as a use case.
Specifications and access
| Specification | Value and source |
|---|---|
| Model type | Open weightsSource |
| Model ID | IndexTeam/Index-Translate-2BSource |
| Model task | Instruction-following text translation with terminology controlSource |
| Access | Public Hugging Face weights for local executionSource |
| Input modalities | Text and translation instructionsSource |
| Output modalities | Translated textSource |
| License | Apache 2.0 according to the Hugging Face model cardSource |
| Release | Public weights released on September 30, 2026Source |
| Languages | Text translation across 150 languagesSource |
| Controls | Can follow terminology, style, and output-format constraintsSource |
| Parameters | 2 billionSource |
Page 1 of 5
Measurements without matching peer values
These measurements have no matching peer values under the same test conditions. Their original values and sources remain available here.
FLORES_minor_pair / off-target rate (Show measurement, test conditions, and source)
- Source value
- 4.2
- Score
- 4.2
- Metric
- Share of outputs in a language other than the requested target language
- Unit
- %
- Category
- translation
- Direction
- Lower is better
- Sample
- 104000
- Source type
- vendor-reported
- Evaluator
- Bilibili IndexTeam technical report, Index-Translate (arXiv v1)
- Status
- active
- Retrieved at
- 2026-10-07
- Methodology
- Vendor-reported result. Exact evaluated checkpoint: IndexTeam/Index-Translate-2B. Final released 2B/9B checkpoint; report setup: greedy text-translation decoding, 16,384-token context. Same FLORES low-resource devtest: 104,000 inputs across 1,040 directions among 62 languages. Lower is better.
instTrans_minor / off-target rate (Show measurement, test conditions, and source)
- Source value
- 3.97
- Score
- 3.97
- Metric
- Share of outputs in a language other than the requested target language
- Unit
- %
- Category
- translation
- Direction
- Lower is better
- Sample
- 2793
- Source type
- vendor-reported
- Evaluator
- Bilibili IndexTeam technical report, Index-Translate (arXiv v1)
- Status
- active
- Retrieved at
- 2026-10-07
- Methodology
- Vendor-reported result. Exact evaluated checkpoint: IndexTeam/Index-Translate-2B. Final released 2B/9B checkpoint; report setup: greedy text-translation decoding, 16,384-token context. Low-resource instruction-translation set: 2,793 tasks from Chinese or English into low-resource languages. Lower is better.
C-Eval / answer accuracy (Show measurement, test conditions, and source)
- Source value
- 42.6
- Score
- 42.6
- Metric
- Answer accuracy
- Unit
- %
- Category
- translation
- Direction
- Higher is better
- Sample
- 12342
- Source type
- vendor-reported
- Evaluator
- Bilibili IndexTeam technical report, Index-Translate (arXiv v1)
- Status
- active
- Retrieved at
- 2026-10-07
- Methodology
- Vendor-reported result. Exact model version corresponding to IndexTeam/Index-Translate-2B. 12,342 labeled test questions across 52 subjects; matched evaluation subset; answer accuracy. The report does not specify a prompt template or decoding parameters for these general-capability runs.
GPQA-Diamond / answer accuracy (Show measurement, test conditions, and source)
- Source value
- 27.4
- Score
- 27.4
- Metric
- Answer accuracy
- Unit
- %
- Category
- translation
- Direction
- Higher is better
- Sample
- 198
- Source type
- vendor-reported
- Evaluator
- Bilibili IndexTeam technical report, Index-Translate (arXiv v1)
- Status
- active
- Retrieved at
- 2026-10-07
- Methodology
- Vendor-reported result. Exact model version corresponding to IndexTeam/Index-Translate-2B. All 198 GPQA-Diamond questions; matched evaluation subset; answer accuracy. The report does not specify a prompt template or decoding parameters for these general-capability runs.
INCLUDE / answer accuracy (Show measurement, test conditions, and source)
- Source value
- 39.1
- Score
- 39.1
- Metric
- Answer accuracy
- Unit
- %
- Category
- translation
- Direction
- Higher is better
- Sample
- 22639
- Source type
- vendor-reported
- Evaluator
- Bilibili IndexTeam technical report, Index-Translate (arXiv v1)
- Status
- active
- Retrieved at
- 2026-10-07
- Methodology
- Vendor-reported result. Exact model version corresponding to IndexTeam/Index-Translate-2B. 22,639 questions in 44 languages; matched subset; sample-level averaging for the multilingual set. The report does not specify a prompt template or decoding parameters for these general-capability runs.
MMMLU / answer accuracy (Show measurement, test conditions, and source)
- Source value
- 42.7
- Score
- 42.7
- Metric
- Answer accuracy
- Unit
- %
- Category
- translation
- Direction
- Higher is better
- Sample
- 4200
- Source type
- vendor-reported
- Evaluator
- Bilibili IndexTeam technical report, Index-Translate (arXiv v1)
- Status
- active
- Retrieved at
- 2026-10-07
- Methodology
- Vendor-reported result. Exact model version corresponding to IndexTeam/Index-Translate-2B. 4,200 questions: 300 per language across 14 languages; sampling seed 42; matched subset. The report does not specify a prompt template or decoding parameters for these general-capability runs.
instTrans / core / subtitle syllable alignment (Show measurement, test conditions, and source)
- Source value
- 46.6
- Score
- 46.6
- Metric
- binary hard-constraint pass rate
- Unit
- %
- Category
- translation
- Direction
- Higher is better
- Sample
- 601
- Source type
- vendor-reported
- Evaluator
- Bilibili IndexTeam technical report, Index-Translate (arXiv v1)
- Status
- active
- Retrieved at
- 2026-10-07
- Methodology
- Vendor-reported result. Exact final released checkpoint IndexTeam/Index-Translate-2B; core instTrans direction subset; GPT-5.6-Sol assessment for the listed constraint. Table 20 reports 601 items for this constraint.
- Context
- Counts are per constraint; a source example may carry more than one constraint.
instTrans / low-resource / subtitle syllable alignment (Show measurement, test conditions, and source)
- Source value
- 72.8
- Score
- 72.8
- Metric
- binary hard-constraint pass rate
- Unit
- %
- Category
- translation
- Direction
- Higher is better
- Sample
- 765
- Source type
- vendor-reported
- Evaluator
- Bilibili IndexTeam technical report, Index-Translate (arXiv v1)
- Status
- active
- Retrieved at
- 2026-10-07
- Methodology
- Vendor-reported result. Exact final released checkpoint IndexTeam/Index-Translate-2B; low-resource instTrans direction subset; GPT-5.6-Sol assessment for the listed constraint. Table 20 reports 765 items for this constraint.
- Context
- Counts are per constraint; a source example may carry more than one constraint.
FLORES-200 / COMET-22 (Show measurement, test conditions, and source)
- Source value
- 0.8655
- Score
- 0.8655
- Metric
- COMET-22
- Unit
- ratio
- Category
- translation
- Direction
- Higher is better
- Sample
- 126000
- Source type
- vendor-reported
- Evaluator
- Bilibili IndexTeam technical report, Index-Translate (arXiv v1)
- Status
- active
- Retrieved at
- 2026-10-07
- Methodology
- Vendor-reported result. Exact evaluated checkpoint: IndexTeam/Index-Translate-2B. Final released checkpoint; the report describes the 2B/9B post-training as a 0.8:0.1:0.1 interpolation of general, instruction, and meme specialists followed by targeted MOPD. Greedy text-translation decoding, 16,384-token context, except for the separately configured WMT26 run. 126,000 directions/items: 420 directions among 21 languages, devtest, 300 per direction.
WMT24++ / COMET-22 (Show measurement, test conditions, and source)
- Source value
- 0.8489
- Score
- 0.8489
- Metric
- COMET-22
- Unit
- ratio
- Category
- translation
- Direction
- Higher is better
- Sample
- 6000
- Source type
- vendor-reported
- Evaluator
- Bilibili IndexTeam technical report, Index-Translate (arXiv v1)
- Status
- active
- Retrieved at
- 2026-10-07
- Methodology
- Vendor-reported result. Exact evaluated checkpoint: IndexTeam/Index-Translate-2B. Final released checkpoint; the report describes the 2B/9B post-training as a 0.8:0.1:0.1 interpolation of general, instruction, and meme specialists followed by targeted MOPD. Greedy text-translation decoding, 16,384-token context, except for the separately configured WMT26 run. English→20 languages, test split, 300 examples per direction.
WMT26 / GPT-5.6-Sol judge (Show measurement, test conditions, and source)
- Source value
- 60.26
- Score
- 60.26
- Metric
- Reference-free judge score, 0–100 points
- Unit
- points
- Category
- translation
- Direction
- Higher is better
- Sample
- 4897
- Source type
- vendor-reported
- Evaluator
- Bilibili IndexTeam technical report, Index-Translate (arXiv v1)
- Status
- active
- Retrieved at
- 2026-10-07
- Methodology
- Vendor-reported result. Exact evaluated checkpoint: IndexTeam/Index-Translate-2B. Final released checkpoint; the report describes the 2B/9B post-training as a 0.8:0.1:0.1 interpolation of general, instruction, and meme specialists followed by targeted MOPD. Greedy text-translation decoding, 16,384-token context, except for the separately configured WMT26 run. 23 directions covering 20 target languages; zero-shot Chinese instructions; 3,584-token context and 1,536-token output limit; GPT-5.6-Sol reference-free judgments averaged over 20 target languages.
instTrans / translation quality (Show measurement, test conditions, and source)
- Source value
- 0.5391
- Score
- 0.5391
- Metric
- GPT-5.6-Sol translation-quality judgment, mean on 0/0.5/1 rubric
- Unit
- ratio
- Category
- translation
- Direction
- Higher is better
- Sample
- 2961
- Source type
- vendor-reported
- Evaluator
- Bilibili IndexTeam technical report, Index-Translate (arXiv v1)
- Status
- active
- Retrieved at
- 2026-10-07
- Methodology
- Vendor-reported result. Exact evaluated checkpoint: IndexTeam/Index-Translate-2B. Final released checkpoint; the report describes the 2B/9B post-training as a 0.8:0.1:0.1 interpolation of general, instruction, and meme specialists followed by targeted MOPD. Greedy text-translation decoding, 16,384-token context, except for the separately configured WMT26 run. Core instruction-translation set: 3,000 Chinese→20-language tasks. Quality uses three GPT-5.6-Sol judge calls per example on a 0/0.5/1 rubric. Valid judgments: 2B 2,961; 9B 2,960; the report does not state a valid-judgment count for the 35B-A3B preview.
Page 1 of 5
Sources and data date
| Type | Evidence and data date |
|---|---|
| Input | Documented modalities. Retrieved October 7, 2026. |
| Output | Documented modalities. Retrieved October 7, 2026. |
| Research status | Research date October 7, 2026. 4 source URLs checked. This documents the inspected sources, not an exhaustive inventory of every publication. |
| Not found in the inspected sources | independent exact-checkpoint reproduction Original wording The reviewed official cards, evaluation repository, and technical report provide vendor-reported results; they do not cite an independent reproduction of this exact checkpoint. |
| Additional source | Bilibili Index-Translate repository (retrieved October 7, 2026) · Bilibili Index-Translate repository · Editorial description reviewed October 7, 2026 |
| Additional source | Bilibili IndexTeam technical report, Index-Translate (arXiv v1) (retrieved October 7, 2026) |
| Additional source | Bilibili Index-Translate official raw benchmark-score CSV (retrieved October 7, 2026) |
| Additional source | Model metadata (retrieved October 7, 2026) |