Language modelOpen weights
Index-Homura-9B
Bilibili Index Team
- Released
- -
- Data date
- October 7, 2026
- Input
- Text, target syllable count, and optional glossary
- Output
- Translated text adjusted toward an approximate syllable count
Index-Homura-9B offers the same syllable control as the 2B variant in a larger checkpoint. Its translation is designed to preserve meaning while approaching a requested syllable count. That target is approximate and does not guarantee the duration of the spoken line.
Specifications and access
| Specification | Value and source |
|---|---|
| Model type | Open weightsSource |
| Model ID | IndexTeam/Index-Homura-9BSource |
| Model task | Text translation with a target syllable countSource |
| Access | Public Hugging Face weights for local executionSource |
| Input modalities | Text, target syllable count, and optional glossarySource |
| Output modalities | Translated text adjusted toward an approximate syllable countSource |
| License | Apache 2.0 according to the Hugging Face model cardSource |
| Syllable control | Adjusts wording toward a target syllable count; the count does not guarantee spoken durationSource |
| Parameters | 9 billionSource |
Page 1 of 2
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.
SandGlass / outputs within ±1 syllable (Show measurement, test conditions, and source)
- Source value
- 74.42
- Score
- 74.42
- Metric
- Share of cases within one syllable of the target
- Unit
- %
- Category
- translation
- Direction
- Higher is better
- Sample
- 3600
- 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 released RL checkpoint IndexTeam/Index-Homura-9B (not the separate SFT ablation). SandGlass: 300 subtitle sentences × 4 target languages × 3 requested length budgets = 3,600 cases. Local model evaluation used greedy decoding and a 512-token output limit. Gemini-2.5-Flash judged translation quality without references on a 0/0.5/1 scale; target counts derive from subtitle duration and language-specific speaking rates, with short/long targets at 0.75×/1.25×.
SandGlass / outputs within 10% of target (Show measurement, test conditions, and source)
- Source value
- 81.92
- Score
- 81.92
- Metric
- Share of cases within 10% of target syllable count
- Unit
- %
- Category
- translation
- Direction
- Higher is better
- Sample
- 3600
- 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 released RL checkpoint IndexTeam/Index-Homura-9B (not the separate SFT ablation). SandGlass: 300 subtitle sentences × 4 target languages × 3 requested length budgets = 3,600 cases. Local model evaluation used greedy decoding and a 512-token output limit. Gemini-2.5-Flash judged translation quality without references on a 0/0.5/1 scale; target counts derive from subtitle duration and language-specific speaking rates, with short/long targets at 0.75×/1.25×.
SandGlass / English / within 10% of target (Show measurement, test conditions, and source)
- Source value
- 78.67
- Score
- 78.67
- Metric
- Share of outputs within 10% of target syllable count
- Unit
- %
- Category
- translation
- Direction
- Higher is better
- Sample
- 900
- 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 released RL checkpoint IndexTeam/Index-Homura-9B; SandGlass target-language subset: English, 900 cases (300 subtitle sentences × three target-length budgets). The report uses a reference-free Gemini-2.5-Flash quality judgment and counts the share within 10% of the target.
SandGlass / Japanese / within 10% of target (Show measurement, test conditions, and source)
- Source value
- 86.67
- Score
- 86.67
- Metric
- Share of outputs within 10% of target syllable count
- Unit
- %
- Category
- translation
- Direction
- Higher is better
- Sample
- 900
- 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 released RL checkpoint IndexTeam/Index-Homura-9B; SandGlass target-language subset: Japanese, 900 cases (300 subtitle sentences × three target-length budgets). The report uses a reference-free Gemini-2.5-Flash quality judgment and counts the share within 10% of the target.
SandGlass / Arabic / within 10% of target (Show measurement, test conditions, and source)
- Source value
- 79.89
- Score
- 79.89
- Metric
- Share of outputs within 10% of target syllable count
- Unit
- %
- Category
- translation
- Direction
- Higher is better
- Sample
- 900
- 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 released RL checkpoint IndexTeam/Index-Homura-9B; SandGlass target-language subset: Arabic, 900 cases (300 subtitle sentences × three target-length budgets). The report uses a reference-free Gemini-2.5-Flash quality judgment and counts the share within 10% of the target.
SandGlass / Spanish / within 10% of target (Show measurement, test conditions, and source)
- Source value
- 82.44
- Score
- 82.44
- Metric
- Share of outputs within 10% of target syllable count
- Unit
- %
- Category
- translation
- Direction
- Higher is better
- Sample
- 900
- 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 released RL checkpoint IndexTeam/Index-Homura-9B; SandGlass target-language subset: Spanish, 900 cases (300 subtitle sentences × three target-length budgets). The report uses a reference-free Gemini-2.5-Flash quality judgment and counts the share within 10% of the target.
SandGlass / outputs within 5% of target (Show measurement, test conditions, and source)
- Source value
- 50
- Score
- 50
- Metric
- Share of cases with absolute syllable-count deviation no greater than 5% of the requested target
- Unit
- %
- Category
- translation
- Direction
- Higher is better
- Sample
- 3600
- Source type
- vendor-reported
- Evaluator
- Bilibili Index-Translate SandGlass benchmark README (pinned Git commit)
- Status
- active
- Retrieved at
- 2026-10-07
- Methodology
- Vendor-reported result. Exact released RL checkpoint IndexTeam/Index-Homura-9B, matched to the SandGlass README row “index-homura” marked “Ours (9B RL)”. The full benchmark covers 300 subtitle sentences × 4 target languages × 3 requested length targets (3,600 cases). The table reports the share whose absolute syllable-count deviation is at most 5% of the requested target. The official model card documents greedy decoding with a 512-token output limit for the local SandGlass evaluations.
- Context
- The leaderboard rounds this percentage to one decimal place and publishes no separate language or target-length breakdown for this threshold.
SandGlass / translation quality (Show measurement, test conditions, and source)
- Source value
- 0.7863
- Score
- 0.7863
- Metric
- Reference-free Gemini-2.5-Flash translation-quality mean
- Unit
- ratio
- Category
- translation
- Direction
- Higher is better
- Sample
- 3600
- 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 released RL checkpoint IndexTeam/Index-Homura-9B (not the separate SFT ablation). SandGlass: 300 subtitle sentences × 4 target languages × 3 requested length budgets = 3,600 cases. Local model evaluation used greedy decoding and a 512-token output limit. Gemini-2.5-Flash judged translation quality without references on a 0/0.5/1 scale; target counts derive from subtitle duration and language-specific speaking rates, with short/long targets at 0.75×/1.25×.
SandGlass / mean relative syllable-count deviation (Show measurement, test conditions, and source)
- Source value
- 0.0693
- Score
- 0.0693
- Metric
- Mean absolute syllable-count error divided by requested target count
- Unit
- ratio
- Category
- translation
- Direction
- Lower is better
- Sample
- 3600
- 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 released RL checkpoint IndexTeam/Index-Homura-9B (not the separate SFT ablation). SandGlass: 300 subtitle sentences × 4 target languages × 3 requested length budgets = 3,600 cases. Local model evaluation used greedy decoding and a 512-token output limit. Gemini-2.5-Flash judged translation quality without references on a 0/0.5/1 scale; target counts derive from subtitle duration and language-specific speaking rates, with short/long targets at 0.75×/1.25×.
SandGlass / official composite score (Show measurement, test conditions, and source)
- Source value
- 0.866
- Score
- 0.866
- Metric
- Vendor-labeled score; formula and scale not stated
- Unit
- No unit provided
- Category
- translation
- Direction
- Higher is better
- Sample
- 3600
- Source type
- vendor-reported
- Evaluator
- Bilibili Index-Translate official model site
- Status
- active
- Retrieved at
- 2026-10-07
- Methodology
- Vendor-reported result. Exact released checkpoint IndexTeam/Index-Homura-9B; official site labels this as a SandGlass composite / syllable-reward result. Its displayed table does not define the score formula or a measurement unit; it is therefore recorded with unit unknown.
- Context
- Do not interpret this as a percentage or as the separate translation-quality score; the official table does not state the formula or unit.
SandGlass / official syllable reward (Show measurement, test conditions, and source)
- Source value
- 0.9457
- Score
- 0.9457
- Metric
- Vendor-labeled score; formula and scale not stated
- Unit
- No unit provided
- Category
- translation
- Direction
- Higher is better
- Sample
- 3600
- Source type
- vendor-reported
- Evaluator
- Bilibili Index-Translate official model site
- Status
- active
- Retrieved at
- 2026-10-07
- Methodology
- Vendor-reported result. Exact released checkpoint IndexTeam/Index-Homura-9B; official site labels this as a SandGlass composite / syllable-reward result. Its displayed table does not define the score formula or a measurement unit; it is therefore recorded with unit unknown.
- Context
- Do not interpret this as a percentage or as the separate translation-quality score; the official table does not state the formula or unit.
SandGlass / English / translation quality (Show measurement, test conditions, and source)
- Source value
- 0.795
- Score
- 0.795
- Metric
- Reference-free Gemini-2.5-Flash translation-quality mean
- Unit
- ratio
- Category
- translation
- Direction
- Higher is better
- Sample
- 900
- 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 released RL checkpoint IndexTeam/Index-Homura-9B; SandGlass target-language subset: English, 900 cases (300 subtitle sentences × three target-length budgets). The report uses a reference-free Gemini-2.5-Flash quality judgment and counts the share within 10% of the target.
Page 1 of 2
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. 5 source URLs checked. This documents the inspected sources, not an exhaustive inventory of every publication. |
| Unresolved | SandGlass control slope Original wording The report publishes a slope score but defines the ideal as approximately 1, so neither a monotone higher-is-better nor lower-is-better direction is correct for this schema; the value is deliberately not encoded as a ranked benchmark. |
| 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 RL 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 model site (retrieved October 7, 2026) |
| Additional source | Bilibili Index-Translate SandGlass benchmark README (pinned Git commit) (retrieved October 7, 2026) |
| Additional source | Model metadata (retrieved October 7, 2026) |