Agregar NOTICE
Browse files
NOTICE
ADDED
|
@@ -0,0 +1,34 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
KBBridge-v3
|
| 2 |
+
Copyright 2026 InstantGMP
|
| 3 |
+
|
| 4 |
+
This product includes software developed from Qwen3.8-27B
|
| 5 |
+
(https://huggingface.co/Qwen/Qwen3.8-27B), Copyright 2026 Alibaba Cloud,
|
| 6 |
+
licensed under the Apache License, Version 2.0.
|
| 7 |
+
|
| 8 |
+
KBBridge-v3 is a MODIFIED derivative of that model. The modifications are:
|
| 9 |
+
|
| 10 |
+
1. Supervised fine-tuning with LoRA (rank 64, alpha 128, all projections),
|
| 11 |
+
merged into the base weights. The adapter was trained on GeneXus source
|
| 12 |
+
code in .gxSource format.
|
| 13 |
+
|
| 14 |
+
2. Quantisation to GGUF Q8_0 and Q4_K_M using llama.cpp.
|
| 15 |
+
|
| 16 |
+
3. The multi-token-prediction head was carried over unchanged from the base
|
| 17 |
+
model, as the LoRA adapter did not target those tensors.
|
| 18 |
+
|
| 19 |
+
4. The vision tower is carried over unchanged from the base model. The GGUF
|
| 20 |
+
builds in this repository are text-only; the base model's own multimodal
|
| 21 |
+
projector applies to them without modification.
|
| 22 |
+
|
| 23 |
+
The base model's vision and MTP weights are redistributed here as part of the
|
| 24 |
+
merged model, unmodified, under the same Apache 2.0 license.
|
| 25 |
+
|
| 26 |
+
You may obtain a copy of the License at
|
| 27 |
+
|
| 28 |
+
http://www.apache.org/licenses/LICENSE-2.0
|
| 29 |
+
|
| 30 |
+
Unless required by applicable law or agreed to in writing, software
|
| 31 |
+
distributed under the License is distributed on an "AS IS" BASIS, WITHOUT
|
| 32 |
+
WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the
|
| 33 |
+
License for the specific language governing permissions and limitations under
|
| 34 |
+
the License.
|