Text Classification
Transformers
Safetensors
English
Chinese
qwen3_5
image-text-to-text
decision-model
system-one
decision-index
lora-merged
Instructions to use PelaAI/KnowLine-4B-Gen5.5 with libraries, inference providers, notebooks, and local apps. Follow these links to get started.
- Libraries
- Transformers
How to use PelaAI/KnowLine-4B-Gen5.5 with Transformers:
# Use a pipeline as a high-level helper from transformers import pipeline pipe = pipeline("text-classification", model="PelaAI/KnowLine-4B-Gen5.5")# pip install -U transformers accelerate # Load model directly from transformers import AutoProcessor, AutoModelForMultimodalLM processor = AutoProcessor.from_pretrained("PelaAI/KnowLine-4B-Gen5.5") model = AutoModelForMultimodalLM.from_pretrained("PelaAI/KnowLine-4B-Gen5.5", device_map="auto") - Notebooks
- Google Colab
- Kaggle
Model cards: remove the KOF section
Browse files- README.md +0 -15
- README.zh.md +0 -10
- SHA256SUMS +2 -2
README.md
CHANGED
|
@@ -53,7 +53,6 @@ Highlights:
|
|
| 53 |
reference hardware); see [INFERENCE.md](INFERENCE.md).
|
| 54 |
- **Images: unchanged.** Requests with images are sent exactly as in Gen5, so Gen5's image results apply; see
|
| 55 |
[Images](#images-zero-shot).
|
| 56 |
-
- **KOF '98 harness: 16-1-1 against StartLux-Decision-4B and 10-6-2 against KnowLine-4B-Gen4** in single-bout mirror matches; see [Game harness](#game-harness-kof-98).
|
| 57 |
|
| 58 |
KnowLine is an independent model. It is not affiliated with, endorsed by, or derived from TypeSafe or Jev.
|
| 59 |
|
|
@@ -201,20 +200,6 @@ Gen5.5 front end (read twice on). Unless noted, the Gen5 numbers were measured a
|
|
| 201 |
Qwen3.5-4B: 59.4); see the [Gen5 card](https://huggingface.co/PelaAI/KnowLine-4B-Gen5#images-zero-shot) for the
|
| 202 |
per-benchmark table.
|
| 203 |
|
| 204 |
-
### Game harness (KOF '98)
|
| 205 |
-
|
| 206 |
-
- **Setup:** single-bout character-mirror matches, 18 games per pair, argmax actions; the same settings as for Gen5.
|
| 207 |
-
- **Result:** 16-1-1 against StartLux-Decision-4B and 10-6-2 against KnowLine-4B-Gen4; 26-7-3 overall, a score of
|
| 208 |
-
0.764 [0.60, 0.87] (Gen5 against the same two opponents: 14-3-1 and 9-9-0).
|
| 209 |
-
- **No games against Jev this time:** our TypeSafe API credits had run out when these games were played, so every Jev
|
| 210 |
-
request failed; those games are excluded and there is no Jev result for Gen5.5.
|
| 211 |
-
- **How the games ended:** against StartLux-Decision-4B, 7 of the 16 wins were by K.O. and 9 at time-out, and the loss
|
| 212 |
-
and the draw came at time-out; against KnowLine-4B-Gen4, 5 of the 10 wins and 3 of the 6 losses were by K.O., and
|
| 213 |
-
both draws at time-out.
|
| 214 |
-
- **Play style:** special_2 50%, special_1 20%, roll forward 10%, heavy punch 8%, heavy kick 7%.
|
| 215 |
-
- **Caveats:** 18 games per pair give wide intervals; this is a measured result in this harness, not general
|
| 216 |
-
fighting-game skill.
|
| 217 |
-
|
| 218 |
### Calibration
|
| 219 |
|
| 220 |
Computed on the 0.3 public suite using the method described by the Decision Index board: each field is scored right or
|
|
|
|
| 53 |
reference hardware); see [INFERENCE.md](INFERENCE.md).
|
| 54 |
- **Images: unchanged.** Requests with images are sent exactly as in Gen5, so Gen5's image results apply; see
|
| 55 |
[Images](#images-zero-shot).
|
|
|
|
| 56 |
|
| 57 |
KnowLine is an independent model. It is not affiliated with, endorsed by, or derived from TypeSafe or Jev.
|
| 58 |
|
|
|
|
| 200 |
Qwen3.5-4B: 59.4); see the [Gen5 card](https://huggingface.co/PelaAI/KnowLine-4B-Gen5#images-zero-shot) for the
|
| 201 |
per-benchmark table.
|
| 202 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 203 |
### Calibration
|
| 204 |
|
| 205 |
Computed on the 0.3 public suite using the method described by the Decision Index board: each field is scored right or
|
README.zh.md
CHANGED
|
@@ -17,7 +17,6 @@
|
|
| 17 |
- **校准:** 权重里折入的温度不变(T = 1.50)。按榜单的方法计算,ECE 约 0.035-0.046(Gen5 为 0.037-0.049),Brier 分数约 0.27-0.30(Gen5 为 0.28-0.31),见[校准](#校准)。
|
| 18 |
- **代价:** 在 Decision Index 套件上,prompt token 约为 Gen5 的 1.56 倍。单张 H20、一次一个请求:中位数 42.1 ms、平均 83 ms,Gen5 为 31.6 ms 和 58.8 ms(我们自己的测量,不是榜单的参考硬件),见 [INFERENCE.md](INFERENCE.md)。
|
| 19 |
- **图片:不变。** 带图片的请求与 Gen5 的发送方式完全相同,所以 Gen5 的图片结果同样适用,见[图片输入](#图片输入零样本)。
|
| 20 |
-
- **拳皇 '98 评测:对 StartLux-Decision-4B 为 16 胜 1 负 1 平,对 KnowLine-4B-Gen4 为 10 胜 6 负 2 平**(单回合同角色镜像对战),见[游戏评测](#游戏评测拳皇-98)。
|
| 21 |
|
| 22 |
KnowLine 是独立开发的模型,与 TypeSafe 或 Jev 没有隶属关系,未获其背书,也不是基于 TypeSafe 或 Jev 开发的。
|
| 23 |
|
|
@@ -141,15 +140,6 @@ curl -s http://127.0.0.1:8080/v1/systemone -H 'Content-Type: application/json' -
|
|
| 141 |
- **与 Gen5 相同:** 读两遍只作用于纯文本请求。带图片的请求与 Gen5 逐字节相同,发给同样的权重,所以 Gen5 的图片结果对 Gen5.5 同样成立,我们没有重跑。
|
| 142 |
- **Gen5 的结果:** 在我们按 Decision Index 0.3.1 视觉榜公开套件复现的评测上得分 66.0(Gen4 为 66.4,原版 Qwen3.5-4B 为 59.4);各基准的明细见 [Gen5 模型卡](https://huggingface.co/PelaAI/KnowLine-4B-Gen5#images-zero-shot)。
|
| 143 |
|
| 144 |
-
### 游戏评测(拳皇 '98)
|
| 145 |
-
|
| 146 |
-
- **设置:** 单回合同角色镜像对战,每对 18 局,按最高概率出招;设置与 Gen5 相同。
|
| 147 |
-
- **结果:** 对 StartLux-Decision-4B 16 胜 1 负 1 平,对 KnowLine-4B-Gen4 10 胜 6 负 2 平;合计 26 胜 7 负 3 平,得分 0.764 [0.60, 0.87](Gen5 对同样两个对手为 14 胜 3 负 1 平和 9 胜 9 负)。
|
| 148 |
-
- **这次没有对 Jev 的对局:** 跑这批对局时我们的 TypeSafe API 额度已经用完,所有 Jev 请求都失败了;这些对局已排除,所以 Gen5.5 没有对 Jev 的结果。
|
| 149 |
-
- **胜负方式:** 对 StartLux-Decision-4B 的 16 胜里 7 局 K.O.、9 局时间到,一场负局和一场平局都是时间到;对 KnowLine-4B-Gen4,10 胜里 5 局 K.O.,6 负里 3 局被 K.O.,两局平局都是时间到。
|
| 150 |
-
- **出招风格:** special_2 50%、special_1 20%、前滚 10%、重拳 8%、重踢 7%。
|
| 151 |
-
- **注意:** 每对 18 局,置信区间较宽;这是在这套评测里测得的结果,不代表一般的格斗游戏能力。
|
| 152 |
-
|
| 153 |
### 校准
|
| 154 |
|
| 155 |
在 0.3 公开套件上按 Decision Index 榜单描述的方法计算:每个字段判对错,置信度取所选选项的概率,各基准等权,分 10 个等宽区间。
|
|
|
|
| 17 |
- **校准:** 权重里折入的温度不变(T = 1.50)。按榜单的方法计算,ECE 约 0.035-0.046(Gen5 为 0.037-0.049),Brier 分数约 0.27-0.30(Gen5 为 0.28-0.31),见[校准](#校准)。
|
| 18 |
- **代价:** 在 Decision Index 套件上,prompt token 约为 Gen5 的 1.56 倍。单张 H20、一次一个请求:中位数 42.1 ms、平均 83 ms,Gen5 为 31.6 ms 和 58.8 ms(我们自己的测量,不是榜单的参考硬件),见 [INFERENCE.md](INFERENCE.md)。
|
| 19 |
- **图片:不变。** 带图片的请求与 Gen5 的发送方式完全相同,所以 Gen5 的图片结果同样适用,见[图片输入](#图片输入零样本)。
|
|
|
|
| 20 |
|
| 21 |
KnowLine 是独立开发的模型,与 TypeSafe 或 Jev 没有隶属关系,未获其背书,也不是基于 TypeSafe 或 Jev 开发的。
|
| 22 |
|
|
|
|
| 140 |
- **与 Gen5 相同:** 读两遍只作用于纯文本请求。带图片的请求与 Gen5 逐字节相同,发给同样的权重,所以 Gen5 的图片结果对 Gen5.5 同样成立,我们没有重跑。
|
| 141 |
- **Gen5 的结果:** 在我们按 Decision Index 0.3.1 视觉榜公开套件复现的评测上得分 66.0(Gen4 为 66.4,原版 Qwen3.5-4B 为 59.4);各基准的明细见 [Gen5 模型卡](https://huggingface.co/PelaAI/KnowLine-4B-Gen5#images-zero-shot)。
|
| 142 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 143 |
### 校准
|
| 144 |
|
| 145 |
在 0.3 公开套件上按 Decision Index 榜单描述的方法计算:每个字段判对错,置信度取所选选项的概率,各基准等权,分 10 个等宽区间。
|
SHA256SUMS
CHANGED
|
@@ -1,7 +1,7 @@
|
|
| 1 |
535a52ce3bca0cd70a77564f413a28bbf7046cd30517abc7bd9525332f002b36 INFERENCE.md
|
| 2 |
50cbab8a892c5f2993b8c7351a99182507472def3b1374558308605d99b86b32 LICENSE
|
| 3 |
-
|
| 4 |
-
|
| 5 |
a4aee8afcf2e0711942cf848899be66016f8d14a889ff9ede07bca099c28f715 chat_template.jinja
|
| 6 |
ddc63e1c717afa86c865bb5e01313d89d72bb53b97ad4a8a03ba8510c0621670 config.json
|
| 7 |
f888421726665e8a84b738eed42a64875aed79de8be7daade851ac8bf4c0cef9 configuration.json
|
|
|
|
| 1 |
535a52ce3bca0cd70a77564f413a28bbf7046cd30517abc7bd9525332f002b36 INFERENCE.md
|
| 2 |
50cbab8a892c5f2993b8c7351a99182507472def3b1374558308605d99b86b32 LICENSE
|
| 3 |
+
e67f63f5724952994612ab49a510c1622c16fa12033e1dadccf519e07783a92b README.md
|
| 4 |
+
0c801278e98411d7ebc0882a4165a20a55e8db413374dbc3dbd14fe447a39502 README.zh.md
|
| 5 |
a4aee8afcf2e0711942cf848899be66016f8d14a889ff9ede07bca099c28f715 chat_template.jinja
|
| 6 |
ddc63e1c717afa86c865bb5e01313d89d72bb53b97ad4a8a03ba8510c0621670 config.json
|
| 7 |
f888421726665e8a84b738eed42a64875aed79de8be7daade851ac8bf4c0cef9 configuration.json
|