Metrics
Every number we publish, with its definition, its source and the date it was read. All of it comes from one file, which you can open yourself.
How to read this page
Each figure is read from a single file, metrics.json, published by the Sango platform. Every entry in it carries what exactly was counted, where the value comes from, the query that produced it and the date it was last read. Values that can be re-read automatically (download counts, licences, dataset sizes) are refreshed when the platform is built; measured values carry the date of the measurement and the exact query to repeat it. Nothing on this page is typed in by hand, and a figure that cannot be sourced is not here.
Open metrics.json, the file every number on this page comes from (opens in a new tab)
Usage
Small numbers, stated exactly. These two are the only usage figures we publish; they come from a dated database measurement, not from analytics.
207
Chat conversations on the Sango platform
Verified against Sango platform database, dated measurement, 10 Oct 2026.
Distinct conversation identifiers in the platform's chat-message store between 30 January 2026 (first stored message) and 10 October 2026, excluding the platform's own test, smoke-test and probe sessions. One conversation is one chat thread, however many messages it holds. Guest and signed-in conversations are both counted. Translations, vocabulary look-ups and learning activities are not chat and are not in this number. Window: 30 Jan 2026 to 10 Oct 2026.
Retrieved 10 Oct 2026
15
Signed-in users who have used the chat
Verified against Sango platform database, dated measurement, 10 Oct 2026.
Distinct signed-in user identifiers with at least one stored chat message between 30 January and 10 October 2026, excluding test sessions and excluding guests (guest messages carry no identity; 580 of 976 real messages are guest messages). This is not the number of registered accounts and not the number of visitors. Window: 30 Jan 2026 to 10 Oct 2026.
Retrieved 10 Oct 2026
Vocabulary
Two counts that must be read together: how many entries the pipeline has verified, and how many a native speaker has confirmed one by one.
589
Verified vocabulary entries, live platform count
Live value, updated 21 min ago.
Entries whose status is 'verified' in the platform's production database, read live from its public statistics endpoint. Pipeline-verified, not native-speaker sign-off. The count falls when an audit withdraws entries.
Retrieved 13 Oct 2026, 01:23 UTC
590
Entries in the published vocabulary dataset
Verified against Hugging Face dataset file, 10 Oct 2026.
Rows in vocabulary.jsonl of the MEYNG/sango-vocabulary dataset on Hugging Face (release v1.5.5). Differs from the live platform count because the dataset is a dated snapshot of the database.
Retrieved 10 Oct 2026
1
Entries individually confirmed by a native speaker
Verified against Hugging Face dataset file, 10 Oct 2026.
Entries in the published dataset whose source field is 'native_speaker', i.e. confirmed one by one by a native Sango speaker. Every other entry is pipeline-verified only. This is the second half of the dual counter: N verified · M native-verified.
Retrieved 10 Oct 2026
360
Sango–French sentence pairs in the published dataset
Verified against Hugging Face dataset file, 10 Oct 2026.
Rows in training_pairs.jsonl of the MEYNG/sango-vocabulary dataset. Extracted from grammar and dictionary reference material; not the fine-tuning corpus (see training_pairs_v2_2).
Retrieved 10 Oct 2026
Downloads
As reported by the Hugging Face API. The 30-day figures are rolling and move within a day; the all-time figures are cumulative since publication.
241
Vocabulary dataset downloads, last 30 days
Verified against Hugging Face API, 10 Oct 2026.
The 'downloads' field of the Hugging Face datasets API for MEYNG/sango-vocabulary. Hugging Face defines it as downloads over the previous 30 days; it is not cumulative and it moves within a single day.
Retrieved 10 Oct 2026
1,085
Vocabulary dataset downloads, all time
Verified against Hugging Face API, 10 Oct 2026.
The 'downloadsAllTime' field of the Hugging Face datasets API for MEYNG/sango-vocabulary: cumulative since the dataset was created on 2 April 2026.
Retrieved 10 Oct 2026
231
Translation model v2-2 downloads, last 30 days
Verified against Hugging Face API, 10 Oct 2026.
The 'downloads' field of the Hugging Face models API for MEYNG/nllb-sango-finetuned-600m-v2-2, the current model. Previous 30 days, not cumulative. The superseded v1 model is a different repository and is not counted here.
Retrieved 10 Oct 2026
297
Translation model v2-2 downloads, all time
Verified against Hugging Face API, 10 Oct 2026.
The 'downloadsAllTime' field of the Hugging Face models API for MEYNG/nllb-sango-finetuned-600m-v2-2: cumulative since the model was created on 17 July 2026.
Retrieved 10 Oct 2026
Model quality
MEYNG/nllb-sango-finetuned-600m-v2-2 against facebook/nllb-200-distilled-600M (untrained), chrF, out of distribution. The direction where we lose is on the same table, at the same size.
| Direction | Test set | This model | Baseline | Difference | Verdict |
|---|---|---|---|---|---|
| Sango → French | FLORES+ devtest · 500 sentences | 31.88 | 31.37 | +0.51 | We win |
| Sango → French | SIB-200 test · 204 sentences | 31.45 | 30.20 | +1.25 | We win |
| French → Sango | FLORES+ devtest · 500 sentences | 32.56 | 33.70 | −1.14 | Baseline wins |
| French → Sango | SIB-200 test · 204 sentences | 34.32 | 36.03 | −1.71 | Baseline wins |
Verified against model card, 6 Sept 2026.
Conditions: beam 4, transformers >= 4.40 < 5.0 (v5 breaks the NLLB tokenizer prefix); neither test set was used for checkpoint selection
Retrieved 6 Sept 2026
chrF is the primary metric; a bare BLEU is never quoted for Sango. We recommend the untrained baseline for French to Sango and publish the number that says so. The scores are those on the model card and were produced with transformers below version 5; a newer transformers changes the NLLB language conditioning and does not reproduce them.
Training data
What the current model was actually trained on, and how much of the public corpus survived cleaning.
45,451
Cleaned unique French–Sango pairs the v2-2 model was trained on
Verified against model card, 6 Sept 2026.
Unique French–Sango sentence pairs remaining after a deterministic cleaning pass over the public NLLB web-mined corpus; trained bidirectionally (90,902 rows). Applies to model v2-2 only. The superseded v1 model was trained on the un-deduplicated corpus ('500,000+ pairs'), a figure that must never be attached to v2-2.
Retrieved 6 Sept 2026
588,039
Pairs advertised by the public French–Sango corpus before cleaning
Verified against model card, 6 Sept 2026.
The public NLLB web-mined French–Sango corpus advertises 588,039 quality-filtered pairs. The cleaning pass found 460,437 duplicate Sango targets, 100,683 rows with leading verse numbers, 29,939 with speaker labels and 12,095 with scripture references; 45,451 unique pairs (7.7%) survive.
Retrieved 6 Sept 2026
The public NLLB web-mined French–Sango corpus advertises 588,039 quality-filtered pairs. The cleaning pass found 460,437 duplicate Sango targets, 100,683 rows with leading verse numbers, 29,939 with speaker labels and 12,095 with scripture references; 45,451 unique pairs (7.7%) survive.
Licences
Read from the registries themselves. The model weights are not open-source and we do not call them that.
CC BY-NC 4.0
Licence of the published translation model weights — MEYNG/nllb-sango-finetuned-600m-v2-2, MEYNG/nllb-sango-finetuned-600m, MEYNG/nllb-sango-finetuned-600m-3ep, MEYNG/nllb-sango-finetuned-600m-v2-3
Inherited from Meta's NLLB-200 base model and not ours to change. Non-commercial: research and evaluation only. It is a published, openly benchmarked model, but it is not an open-source licence and it must not be described as one.
Verified against Hugging Face API, 10 Oct 2026.
GET https://huggingface.co/api/models/MEYNG/nllb-sango-finetuned-600m-v2-2 → cardData.license (cc-by-nc-4.0)
Retrieved 10 Oct 2026
CC BY-SA 4.0
Licence of the published vocabulary dataset — MEYNG/sango-vocabulary
Chosen by MEYNG. Commercial use allowed; derivatives must be shared under the same licence. The grant covers MEYNG's curation; 316 of 590 entries are derived from the NLLB corpus and that upstream chain is under review, as the dataset's own DATASHEET states.
Verified against Hugging Face API, 10 Oct 2026.
GET https://huggingface.co/api/datasets/MEYNG/sango-vocabulary → cardData.license (cc-by-sa-4.0)
Retrieved 10 Oct 2026
MIT
Licence of the sango-nlp toolkit on npm — sango-nlp (npm), github.com/meyng-hub/sango-nlp
The published npm package sango-nlp (unscoped; there is no scoped variant of it) is MIT-licensed. Its source is public at github.com/meyng-hub/sango-nlp since 10 October 2026.
Verified against npm registry, 10 Oct 2026.
GET https://registry.npmjs.org/sango-nlp → license
Retrieved 10 Oct 2026
MIT
Licence of the public audio-pipeline repository — github.com/meyng-hub/sango-audio-pipeline
One of the two public code repositories under the meyng-hub GitHub organisation that belong to this project (the other is sango-nlp, the npm toolkit's source). The SangoAI application repository is private and is not cited anywhere.
Verified against GitHub, 10 Oct 2026.
GET https://api.github.com/repos/meyng-hub/sango-audio-pipeline → license.spdx_id, visibility
Retrieved 10 Oct 2026
Company
Registry facts about MEYNG SARL, the commercial entity.
979361052
MEYNG SARL company registration (SIREN)
Verified against French company registry, 22 Sept 2026.
French company registry number of MEYNG SARL, RCS Paris, registered 30 August 2023. MEYNG SARL is the commercial vehicle; the HCW association (SIREN 841 629 157) is a different legal person.
Retrieved 22 Sept 2026
1224310
UN Global Marketplace supplier registration
Verified against certificate on request, 10 Oct 2026.
MEYNG's supplier registration number on the United Nations Global Marketplace (UNGM). UNGM does not expose supplier records publicly; the registration certificate is available on request at contact@meyng.com.
Retrieved 10 Oct 2026
Withdrawn
Figures we used to publish and no longer do. They stay here, struck through, with the reason.
72
Translations served
This claim has been withdrawn: The figure came from a quota table that only recorded 72 metered requests between 13 and 22 April 2026, yet it was labelled as an all-time count on both sites. The platform has no all-time translation counter today; the figure is withdrawn until one exists.
The figure came from a quota table that only recorded 72 metered requests between 13 and 22 April 2026, yet it was labelled as an all-time count on both sites. The platform has no all-time translation counter today; the figure is withdrawn until one exists.
1,018
Sessions
This claim has been withdrawn: The counter read the row count of the chat-message store, where one row is one message, not one session: 1,018 rows were 514 user messages and 504 replies across 224 conversations. Replaced by chat_conversations and chat_identified_users, which name what they count.
The counter read the row count of the chat-message store, where one row is one message, not one session: 1,018 rows were 514 user messages and 504 replies across 224 conversations. Replaced by chat_conversations and chat_identified_users, which name what they count.
5 million / 5M+ / 5.5M
Sango speakers
This claim has been withdrawn: No source separating first- and second-language speakers was ever cited, and three different figures circulated. Replaced by the qualitative statement 'Sango, the national language of the Central African Republic' until a citable source exists.
No source separating first- and second-language speakers was ever cited, and three different figures circulated. Replaced by the qualitative statement 'Sango, the national language of the Central African Republic' until a citable source exists.
66M+ speakers across 5 languages (Lingala 25M, Wolof 10M, Bambara 14M, Kirundi 12M)
Language expansion roadmap
This claim has been withdrawn: Only Sango is live. The roadmap named languages that are not in production and speaker counts with no source, and the two sites disagreed on which languages were listed. Removed; MEYNG publishes what is live.
Only Sango is live. The roadmap named languages that are not in production and speaker counts with no source, and the two sites disagreed on which languages were listed. Removed; MEYNG publishes what is live.
To check any figure, open the file and compare: https://sangoai.app/metrics.json (opens in a new tab)