snipedbywifi/soren-models

Soren models

Small models that Soren (Software Reasoning Navigator) can download to build its cards on your own machine. Soren fetches these files from its Settings page, checks each one against a pinned SHA-256, and never downloads anything until you press the button.

File	What it is	Size
soren-lfm3-r8.gguf	A rank-8 LoRA adapter for LFM3 1.7B that writes Soren's calculator cards.	64 MB
card-classifier/model.int8.onnx	An int8 ONNX encoder that decides whether a message deserves a card, and of which kind.	34 MB
card-classifier/tokenizer.json, card-classifier/config.json	Its tokenizer and thresholds.	under 1 MB

The LFM3 1.7B weights themselves are not here. Soren downloads them from LiquidAI/LFM3-1.7B-GGUF.

The adapter

soren-lfm3-r8.gguf is applied on top of LFM3-1.7B-Q4_K_M.gguf through llama-server --lora. It needs llama.cpp build 11960 or newer, because older builds cannot load LFM3.

Built with LFM3. The adapter is a derivative of Liquid AI's LFM3 1.7B and is distributed under the LFM Open License v1.0, a copy of which is in the LFM3 repository. That licence allows use, modification and redistribution, with commercial use limited to organisations under US$10 million in annual revenue. The files in this repository were changed from the original model; they are an adapter and not the original weights.

The classifier

A fine-tune of BAAI/bge-small-en-v1.5 (33M parameters) that predicts one of none, compare, weigh, rank, place, steps or draft for a chat message, with confidence thresholds for how readily a card is offered. It is served by scripts/card_classifier_server.py in the Soren repository. The training data is not published.

Licences

• soren-lfm3-r8.gguf: LFM Open License v1.0, as above.
• card-classifier/*: derived from BAAI/bge-small-en-v1.5, which carries its own licence (check its model card). Soren's own additions are released under the same terms as Soren, Apache 2.0.

Checksums (SHA-256):

bbb2174179d854c983769579f552cb2a69f21ff8a50ea3a453930729e7a8c61f  soren-lfm3-r8.gguf
238b4ba15e97a027a9b11b94e0b30d5ab8f21897bbbb3fe4183dd831cf9d6d5b  card-classifier/model.int8.onnx
c3bbeac7c7669ec478d44bac871031a47b39935533558d252d5b7699e7d4d457  card-classifier/tokenizer.json
0722d2adc6f95c20d03ece451efc196372b9d5da5eca7fbdc1c38de7265978f3  card-classifier/config.json