Downloads · 30 days
50
21% of all-time downloads
YFC-112358/Qwen3.8-27B-Della-Carnice-Ostrich-Salience-Glimmer-v5c
Qwen3.8-27B-Della-Carnice-Ostrich-Salience-Glimmer-v5c is a image-text-to-text model from YFC-112358. Use it for the image-text-to-text task on the model card, and read the license before you ship it in a product. It is set up for transformers.
基于 Qwen/Qwen3.8-27B 的单阶段 dellalinear 融合,四个源同时入场。
Downloads · 30 days
50
21% of all-time downloads
All-time downloads
237
Public
Parameters
27.8B
55.6 GB on disk
Likes
2
Public
Click a slice to open those files.
.safetensors55.6 GB · 100%
From the Hugging Face model README
基于 Qwen/Qwen3.8-27B 的单阶段 della_linear 融合,四个源同时入场。
base_model: Qwen/Qwen3.8-27B
merge_method: della_linear
dtype: bfloat16
tokenizer:
source: base
parameters:
int8_mask: true
normalize: true
lambda: 1.0
models:
- model: kai-os/Carnice-V3
parameters:
weight: 0.54
density: 0.99
epsilon: 0.08
- model: etemiz/Ostrich-27B-Qwen3.8-260815
parameters:
weight: 0.004
density: 0.5
epsilon: 0.32
- model: vectionlabs/Salience-27B-R5
parameters:
weight: 0.3
density: 0.6
epsilon: 0.32
- model: YFC-112358/Qwen3.8-27B-TM-Gemma4-Glimmer-v3
parameters:
weight: 0.65
density: 0.99
epsilon: 0.08
| 源 | weight | density | epsilon | 名义占比 | 实测占比 |
|---|---|---|---|---|---|
kai-os/Carnice-V3 | 0.54 | 0.99 | 0.08 | 36.1% | 3.9% |
etemiz/Ostrich-27B-Qwen3.8-260815 | 0.004 | 0.5 | 0.32 | 0.3% | 48.8% |
vectionlabs/Salience-27B-R5 | 0.3 | 0.6 | 0.32 | 20.1% | 17.2% |
YFC-112358/Qwen3.8-27B-TM-Gemma4-Glimmer-v3 | 0.65 | 0.99 | 0.08 | 43.5% | 0.0% |
‖λ·δ_mix‖ / ‖base‖ 中位数:0.0006(全量构建)20260827,按张量名 + 行区间氾生 ⇒ 断点续传逐位可重现normalize: true ⇒ 每张张量的增量除以当场源的权重之和(四源齐全时 Σw = 1.494)。所以名义 weight 只決定相对比例,不決定绝对强度。int8_mask: true 在本实现里是空操作(它只是 mergekit 的显存优化开关,不改变数值)。della_linear 不做符号选举,所以不存在“少数源被多数源否决”这回事;剪完直接加权相加。想要符号选举把 Cell 1 的 CONSENSUS 改成 True。Qwen/Qwen3.8-27B(tokenizer.source = base),四个源自己加的 special token 不会进来。from transformers import AutoModelForCausalLM, AutoTokenizer
m = AutoModelForCausalLM.from_pretrained("YFC-112358/Qwen3.8-27B-Della-Carnice-Ostrich-Salience-Glimmer-v5c", torch_dtype="bfloat16", device_map="auto")
tk = AutoTokenizer.from_pretrained("YFC-112358/Qwen3.8-27B-Della-Carnice-Ostrich-Salience-Glimmer-v5c")
由 Qwen38_v5_della_linear.ipynb 逐张量流式构建(单机、不落全量磁盘)。