Transparency — our numbers, with their sources
The figures we cite in applications and partnerships, each from a named source with its date and its definition. No number on this page is typed by hand: live values are fetched in your browser, measured values are read from metrics.json.
Live
Live figures load in your browser — open this page with JavaScript enabled to see current numbers
Measured
Read from metrics.json — the single file every public MEYNG number comes from — with the date of the measurement and what exactly was counted.
What is counted
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.
What is counted
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.
What is counted
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.
Withdrawn
Figures this page used to show and no longer does. They stay listed, with the reason.
Withdrawn 10 Oct 2026. 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 The platform has no all-time translation counter today; the figure is withdrawn until one exists.
Withdrawn 10 Oct 2026. 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.
What “verified” means
Entries are verified by our review pipeline against source dictionaries and the production corpus. Individual native-speaker sign-off is recorded per entry — today only one entry carries it, and closing that gap is the core of our current audio and elicitation work. We say “pipeline-verified”, not “native-speaker verified”.
Counts move in both directions
When an audit finds entries whose published meaning the corpus contradicts, we withdraw them — the verified count fell in July 2026 for exactly that reason. A figure you saw here last month may be higher than today's. That is the system working.
Where we stand on the data-to-AI value chain
Five stages, one honest sentence each, so a partner or funder can place us in ten seconds. Every number below is read from metrics.json, as of 12 Nov 2026.
- 1 · Data collection
The published dataset holds 360 Sango–French sentence pairs extracted from reference grammars and dictionaries, each with its provenance recorded; no speech data has been published yet, and no machine-generated Sango enters the corpus as attested.
- 2 · CurationBottleneck
590 vocabulary entries are curated by our review pipeline, but only 1 has been confirmed one by one by a native speaker: reviewer throughput in Bangui is the stage that holds every other stage back.
- 3 · Model trainingProven method
No model ships unless it meets or beats the untrained baseline on out-of-distribution chrF (FLORES+ devtest and SIB-200 test); the current fine-tune passes Sango→French (31.88 vs 31.37 chrF on FLORES+) and fails French→Sango (32.56 vs 33.70), and we publish both results.
- 4 · Deployment
The translation platform and the WhatsApp learning bot are built for low-bandwidth mobile use; because French→Sango did not pass the gate, no raw French→Sango output from the fine-tune ships, and its non-commercial licence keeps it out of any paid service.
- 5 · Re-use
The vocabulary dataset is released under CC BY-SA 4.0 and the model weights under CC BY-NC 4.0, with download counts shown above; a larger inherited parallel corpus stays unreleased until its licence chain is cleared.
Who runs what
eNdara is the educational program of the HCW association (est. 2009). The platform and WhatsApp bot are operated by MEYNG and powered by the SangoAI API.
Check us yourself
- metrics.json — every public number, with its definition, source and date
- All our metrics, explained — meyng.com/metrics
- Vocabulary dataset (CC-BY-SA-4.0) — HuggingFace
- Fine-tuned model with its full evaluation, including where it loses: French→Sango is still below the NLLB-200 baseline out-of-distribution — model card
- Audio recording pipeline (MIT) — GitHub