Data as of October 1, 2026 · How the score is built
XLM-R large (550M) (Belebele Translate-Train-All)
Decision readingXLM-R large (550M) (Belebele Translate-Train-All) is tracked, but not publicly ranked yet. The original benchmark tables below include the published metrics for this evaluated configuration. Their distinct protocols remain outside general model rankings.
XLM-R large (550M) (Belebele Translate-Train-All) will be repriced, updated or retired. Get each notice with its source and date. Follow model changes
Share or export
Original benchmark results
Every published metric for this configuration, with source precision and evaluation limits. Download results (JSON) · Numeric metrics (CSV)
Paper summary across 122 language variants
Belebele reports multiple-choice accuracy over 122 language variants, with 900 questions per variant. Few-shot, zero-shot, English fine-tuning, translated training, and translated tests keep separate configurations. All published per-language values are included. Summary and appendix averages sometimes disagree: GPT-3.5 is 51.1 versus 50.6, and the 91-language Llama-2-chat subset is 44.1 versus 44.0. Table 2 labels a Llama 1 row 70B while Table 7 identifies 65B; the profile uses the appendix identity and retains the conflicting label. BLOOMZ used FLORES material during training, which can advantage its evaluation.
Source vocabulary and parameter sizes are metadata, not scores. Source summary averages remain unchanged despite appendix discrepancies.
Published source · Full Belebele results
The Belebele Benchmark: a Parallel Reading Comprehension Dataset in 122 Language Variants — Lucas Bandarkar and colleagues. CC BY SA 4.0. Copyright 2024 the authors. Numeric results transcribed and reformatted; evaluation notes and model links added by BenchLM. Published numeric precision is retained.
| Model | Size or variant | Vocabulary size | Average accuracy (%) | Language share ≥50% accuracy (%) | Language share ≥70% accuracy (%) | English accuracy (%) | Non-English average (%) |
|---|---|---|---|---|---|---|---|
| XLM-R | large (550M) | 250K | 58.9 | 69.7% | 36.1% | 78.7 | 58.8 |
All 122 language results: multilingual encoders
Belebele reports multiple-choice accuracy over 122 language variants, with 900 questions per variant. Few-shot, zero-shot, English fine-tuning, translated training, and translated tests keep separate configurations. All published per-language values are included. Summary and appendix averages sometimes disagree: GPT-3.5 is 51.1 versus 50.6, and the 91-language Llama-2-chat subset is 44.1 versus 44.0. Table 2 labels a Llama 1 row 70B while Table 7 identifies 65B; the profile uses the appendix identity and retains the conflicting label. BLOOMZ used FLORES material during training, which can advantage its evaluation.
Accuracy values are percentages. PCT rows are the percentage of language variants meeting an accuracy threshold; summary and appendix PCT values differ and are both retained.
Published source · Full Belebele results
The Belebele Benchmark: a Parallel Reading Comprehension Dataset in 122 Language Variants — Lucas Bandarkar and colleagues. CC BY SA 4.0. Copyright 2024 the authors. Numeric results transcribed and reformatted; evaluation notes and model links added by BenchLM. Published numeric precision is retained.
| Language or summary metric | XLM-R large (550M) (Belebele Translate-Train-All) |
|---|---|
| AVG | 58.9 |
| PCT Above 50 | 69.7% |
| PCT Above 70 | 36.4% |
| eng_Latn | 78.7 |
| acm_Arab | 59.2 |
| afr_Latn | 74.3 |
| als_Latn | 71.4 |
| amh_Ethi | 60.7 |
| apc_Arab | 61.9 |
| arb_Arab | 74.3 |
| arb_Latn | 30.6 |
| ars_Arab | 65.9 |
| ary_Arab | 52.6 |
| arz_Arab | 66.1 |
| asm_Beng | 56.9 |
| azj_Latn | 65.1 |
| bam_Latn | 36.9 |
| ben_Beng | 63.7 |
| ben_Latn | 48.1 |
| bod_Tibt | 36.9 |
| bul_Cyrl | 74.2 |
| cat_Latn | 74.7 |
| ceb_Latn | 50.7 |
| ces_Latn | 74.4 |
| ckb_Arab | 36.9 |
| dan_Latn | 74.7 |
| deu_Latn | 76.7 |
| ell_Grek | 73.0 |
| est_Latn | 70.4 |
| eus_Latn | 70.3 |
| fin_Latn | 74.9 |
| fra_Latn | 75.6 |
| fuv_Latn | 31.1 |
| gaz_Latn | 43.3 |
| grn_Latn | 41.9 |
| guj_Gujr | 63.1 |
| hat_Latn | 39.8 |
| hau_Latn | 53.0 |
| heb_Hebr | 70.6 |
| hin_Deva | 63.4 |
| hin_Latn | 58.9 |
| hrv_Latn | 74.0 |
| hun_Latn | 72.8 |
| hye_Armn | 64.7 |
| ibo_Latn | 32.2 |
| ilo_Latn | 39.7 |
| ind_Latn | 70.4 |
| isl_Latn | 69.0 |
| ita_Latn | 73.3 |
| jav_Latn | 66.8 |
| jpn_Jpan | 71.0 |
| kac_Latn | 33.3 |
| kan_Knda | 69.1 |
| kat_Geor | 67.4 |
| kaz_Cyrl | 64.7 |
| kea_Latn | 48.1 |
| khk_Cyrl | 64.2 |
| khm_Khmr | 63.8 |
| kin_Latn | 38.6 |
| kir_Cyrl | 67.7 |
| kor_Hang | 74.8 |
| lao_Laoo | 63.0 |
| lin_Latn | 34.4 |
| lit_Latn | 72.0 |
| lug_Latn | 34.7 |
| luo_Latn | 33.2 |
| lvs_Latn | 73.0 |
| mal_Mlym | 67.1 |
| mar_Deva | 67.2 |
| mkd_Cyrl | 72.8 |
| mlt_Latn | 42.7 |
| mri_Latn | 34.0 |
| mya_Mymr | 62.9 |
| nld_Latn | 72.8 |
| nob_Latn | 74.2 |
| npi_Deva | 62.7 |
| npi_Latn | 34.8 |
| nso_Latn | 34.7 |
| nya_Latn | 30.8 |
| ory_Orya | 63.9 |
| pan_Guru | 62.0 |
| pbt_Arab | 61.1 |
| pes_Arab | 72.0 |
| plt_Latn | 58.1 |
| pol_Latn | 72.7 |
| por_Latn | 74.0 |
| ron_Latn | 74.8 |
| rus_Cyrl | 77.1 |
| shn_Mymr | 27.0 |
| sin_Latn | 37.3 |
| sin_Sinh | 68.6 |
| slk_Latn | 73.0 |
| slv_Latn | 73.9 |
| sna_Latn | 35.9 |
| snd_Arab | 61.3 |
| som_Latn | 50.7 |
| sot_Latn | 32.7 |
| spa_Latn | 76.4 |
| srp_Cyrl | 75.9 |
| ssw_Latn | 38.9 |
| sun_Latn | 59.4 |
| swe_Latn | 75.1 |
| swh_Latn | 68.7 |
| tam_Taml | 65.3 |
| tel_Telu | 61.1 |
| tgk_Cyrl | 39.6 |
| tgl_Latn | 70.0 |
| tha_Thai | 70.1 |
| tir_Ethi | 37.7 |
| tsn_Latn | 34.3 |
| tso_Latn | 37.1 |
| tur_Latn | 72.0 |
| ukr_Cyrl | 75.0 |
| urd_Arab | 66.3 |
| urd_Latn | 48.4 |
| uzn_Latn | 70.2 |
| vie_Latn | 71.4 |
| war_Latn | 46.6 |
| wol_Latn | 32.2 |
| xho_Latn | 45.4 |
| yor_Latn | 27.9 |
| zho_Hans | 74.8 |
| zho_Hant | 71.3 |
| zsm_Latn | 72.2 |
| zul_Latn | 44.1 |
Multilingual encoder training settings
Belebele reports multiple-choice accuracy over 122 language variants, with 900 questions per variant. Few-shot, zero-shot, English fine-tuning, translated training, and translated tests keep separate configurations. All published per-language values are included. Summary and appendix averages sometimes disagree: GPT-3.5 is 51.1 versus 50.6, and the 91-language Llama-2-chat subset is 44.1 versus 44.0. Table 2 labels a Llama 1 row 70B while Table 7 identifies 65B; the profile uses the appendix identity and retains the conflicting label. BLOOMZ used FLORES material during training, which can advantage its evaluation.
These are training settings, not benchmark scores. XLM-R Translate-Train-All learning rate is printed as 3-6 in the source; no exponent is inferred.
Published source · Full Belebele results
The Belebele Benchmark: a Parallel Reading Comprehension Dataset in 122 Language Variants — Lucas Bandarkar and colleagues. CC BY SA 4.0. Copyright 2024 the authors. Numeric results transcribed and reformatted; evaluation notes and model links added by BenchLM. Published numeric precision is retained.
| Training setting | XLM-R large (550M) (Belebele Translate-Train-All) |
|---|---|
| epochs | 1 |
| training set size | 650k |
| learning rate | 3e-6 |
| weight decay | 0.001 |
| batch size | 64 |
Lineage
The sequence follows explicit supersedes links. Each score is estimated for that model; a relative can inform a sparse estimate but never sets a floor, so a newer release can score below an earlier one. Scores and prices remain blank when the corresponding public row or first-party rate is unavailable.
Release date not sourced · you are here
XLM-R large (550M) (Belebele Translate-Train-All)Not publicly ranked · Price not listed
Benchmark-system
Spec sheet
Each documented value carries its source. Missing fields stay visible as not sourced or not published, rather than disappearing from the page.
- API model ID
- Not publishedProvider pricing
- Context window
- Not published
- Maximum output
- Not sourced yet
- Knowledge cutoff
- Not sourced yet
- Input modalities
- Not sourced yet
- Output modalities
- Not sourced yet
- Parameters
- Not sourced yet
- Availability
- Not sourced yet
- Cloud regions
- Not tracked yet
- Lifecycle
- Tracked
- API capabilities
- Tool calling, structured outputs, and batch support are not tracked yet
- Prompt caching
- Not documented in the pricing recordProvider pricing
- Self-host
- Weights are not published
- Rate limits
- Not tracked yet
How to read this profile
The visual layer above carries the decisions. These notes preserve the model, ranking, coverage, and family context behind the numbers.
The original benchmark tables show this evaluated configuration’s published results. Training choices, prompts, response splits, and metrics remain attached to each table. The profile stays outside general model rankings.
XLM-R large (550M) (Belebele Translate-Train-All) is a research system model. No explicit reasoning mode is documented in this profile.
Evaluated system: XLM-R large (550M) (Belebele Translate-Train-All). The linked benchmark-owner report supplies the full numeric results and protocol. This profile represents that evaluated configuration; it does not transfer results to a base model or another prompt. Context limit, current API tariff, and release date are not established by this evaluation.
The original benchmark results appear in the source tables above; they remain separate from weighted benchmark slots.
Last updated October 1, 2026. Runtime fields remain blank until a sourced snapshot exists.
Questions
How does XLM-R large (550M) (Belebele Translate-Train-All) perform overall in AI benchmarks?
This configuration has original benchmark results in the source tables on this profile. Every published metric retains its evaluated configuration, source precision, and protocol. These results stay outside general model rankings; they do not produce a public overall score or transfer to a base model with different settings.
Does XLM-R large (550M) (Belebele Translate-Train-All) have full benchmark coverage on BenchLM?
This configuration has results in the original benchmark tables on this profile. Those tables preserve every imported source metric, but they do not fill the site’s general scoring slots. Other benchmarks remain unmeasured, and compatible configurations are required before comparing results across different reports.
What is the context window size of XLM-R large (550M) (Belebele Translate-Train-All)?
XLM-R large (550M) (Belebele Translate-Train-All)'s context window is not documented in a source tied to this exact model yet. The profile leaves the value unavailable instead of borrowing a limit from an earlier family member or an unverified route. Maximum output length remains a separate field.