Unchanted/omi-models

Omi models

Exact model files the Omi Android app downloads. The app checks each file's size and sha256 before it uses it.

Span reader

span-ft-tiny-q8_0.gguf with span_heads.json and span_heads.bin.

• What it does: it reads up to 5 short message rows that are already on screen and marks the words in one row that answer the question. When no row holds the answer, it marks nothing. It never writes words of its own.

• Base model: deepset/tinyroberta-squad2 (82M, CC BY 4.0).

• Tuning: one pass over 18,000 rows of open chat and question data:

  • SQuAD 2.0 (CC BY-SA 4.0);
  • Molweni (Apache 2.0);
  • QAConv (CC0).

  No private or personal data was used.

• Files:

  • the encoder exported to GGUF at q8_0 and run with llama.cpp (--embeddings --pooling none);
  • the start and end head (span_heads.json describes it, span_heads.bin holds its weights).

• Status: a first version, used in shadow while it is tested.

File	sha256
span-ft-tiny-q8_0.gguf	c11f51e44af49e2e3f48260370c4b32c615099583b37e63e33a80027750ed905
span_heads.json	75dba1f656cb74bd72df91be77ca4a0d255e6960eeb520b0bd8abaabf32e45c1
span_heads.bin	4bc194c0d95b6a70cf3de3171d735e061f1c0120dbaeb86bade2a5bffadbe10d

Licence and credit

The weights are derived from tinyroberta-squad2 (deepset, CC BY 4.0) and tuned partly on SQuAD 2.0 (Rajpurkar et al., CC BY-SA 4.0). They are shared under CC BY-SA 4.0.