card: the whole pipeline incl. the VQ runtime is open (Apache-2.0); contact info@thinletter.io
Browse files
README.md
CHANGED
|
@@ -79,6 +79,14 @@ report: https://github.com/rosecky/embedding-quantization-public.
|
|
| 79 |
|
| 80 |
sha256: `fe1011d5ec5cff995cb687f42afc47edcaf68cc8296dda11aeb7bc9640b8e295` (Q4_K_M + q4_0), `b977fd649e0750fa8092970ff8bb8d8f955e684d9cbbef055bbe89061af38fdc` (Q3_K + q4_0).
|
| 81 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 82 |
## Licence
|
| 83 |
|
| 84 |
Weights derive from `BAAI/bge-m3` (MIT) and are released under MIT. Calibration text: Czech Wikipedia (CC BY-SA;
|
|
|
|
| 79 |
|
| 80 |
sha256: `fe1011d5ec5cff995cb687f42afc47edcaf68cc8296dda11aeb7bc9640b8e295` (Q4_K_M + q4_0), `b977fd649e0750fa8092970ff8bb8d8f955e684d9cbbef055bbe89061af38fdc` (Q3_K + q4_0).
|
| 81 |
|
| 82 |
+
## Everything behind these files is open (2026-09-16)
|
| 83 |
+
|
| 84 |
+
The complete pipeline is public under Apache-2.0 in https://github.com/rosecky/embedding-quantization-public: the verification recipe, the GPTQ exporter onto the llama.cpp
|
| 85 |
+
K-quant grids, the `llama-quantize` recipes of these files, and since 2026-09-16 also the vector-quantisation side (quantiser,
|
| 86 |
+
`.vqw` compiler, structured rotation, post-pass and the WebGPU runtime, `client/vqweb`). The public demo at
|
| 87 |
+
https://thinletter.io/demo runs both the scalar files (wllama) and the vector-quantised containers. Questions and index-specific
|
| 88 |
+
requests: info@thinletter.io, or the issues of the repository.
|
| 89 |
+
|
| 90 |
## Licence
|
| 91 |
|
| 92 |
Weights derive from `BAAI/bge-m3` (MIT) and are released under MIT. Calibration text: Czech Wikipedia (CC BY-SA;
|