Model files that the DataFlow desktop app downloads when someone asks for them. Each is checked against the SHA-256 below before the app uses it, and runs only on the user's own computer.
DataFlow checks field survey data while the team is still in the field: it pulls interviews from SurveyLoopr, SurveyCTO or ODK Central, runs data-quality checks on every sync, and gives one ranked list of problems to fix. These small models help it on the computer itself, with no AI service involved: one finds personal data in answers, and one suggests which existing option an "other, specify" answer means. Download DataFlow · Security and privacy
| File | decider-0.8b-Q8_0.gguf (811,843,136 bytes) |
| SHA-256 | 6cba195228e7c01196cebfdd758c70a6cbf4672fe3247469b2a53cd343aaa7b7 |
| Source | Mapika/decider-0.8b, revision a0a01d6f8135298f400a8c856b355793012ae971 |
| Base model | Qwen/Qwen3.5-0.8B-Base |
| Licence | Apache License 2.0 (LICENSE-decider), as the source and base model |
Changes: converted to GGUF with 8-bit (Q8_0) weights, without the multi-token prediction
layer the checkpoint's configuration names but doesn't ship. The weights are otherwise unchanged.
Made with convert_hf_to_gguf.py --no-mtp --outtype q8_0 from the llama.cpp version bundled in
the llama-cpp-sys-2 0.1.160 Rust crate.
| File | multilingual-e5-small-f16.gguf (242,360,416 bytes) |
| SHA-256 | 527845993140bc735563f7ea3b09952fd460afaebd9bb11e2f55ba54a71120c5 |
| Source | intfloat/multilingual-e5-small, revision 614241f622f53c4eeff9890bdc4f31cfecc418b3 |
| Licence | MIT (LICENSE-multilingual-e5-small) |
Changes: converted to GGUF with 16-bit weights, otherwise unchanged. The model has BERT
weights with an XLM-RoBERTa (sentencepiece) tokenizer, which the stock converter doesn't read
together; convert/convert_e5.py runs convert_hf_to_gguf.py --outtype f16
with the XLM-RoBERTa tokenizer reader (same llama.cpp version as above).
DataFlow is made by ARCED International.