CountingSheep commited on
Commit
f7ad89f
·
verified ·
1 Parent(s): 80e933d

Card: the second generation of the 4B model is CountingSheep/vev-4b-v2

Browse files
Files changed (1) hide show
  1. README.md +11 -13
README.md CHANGED
@@ -43,15 +43,14 @@ Code, API and full results: [Xiaooolong/vev](https://github.com/Xiaooolong/vev).
43
 
44
  <img src="https://raw.githubusercontent.com/Xiaooolong/vev/main/docs/doom.webp" alt="vev-4b playing Doom in real time: the game on the left, the eight slice probabilities and the chosen action on the right.">
45
 
46
- vev-4b 0.1.0 playing Doom in real time through a small harness: 8 yes/no questions per look, one request. Try
47
- vev-4b in the browser: [CountingSheep/vev](https://huggingface.co/spaces/CountingSheep/vev) (runs 0.1.0).
48
 
49
  This repository holds the LoRA adapter of [CountingSheep/vev-4b](https://huggingface.co/CountingSheep/vev-4b) (130
50
  MB). `vev serve` applies it to [Qwen/Qwen3.5-4B](https://huggingface.co/Qwen/Qwen3.5-4B) at load time, so it is the
51
- smaller download if you already have the base model. It holds the 0.1.0 adapter only: there is no 0.2.0 adapter.
52
- `vev-4b` 0.2.0 is published as merged weights in
53
- [CountingSheep/vev-4b](https://huggingface.co/CountingSheep/vev-4b), because its last training stage trains a
54
- second adapter on the merged weights of the stage before. This card describes 0.1.0.
55
 
56
  <img src="https://raw.githubusercontent.com/Xiaooolong/vev/main/docs/checkout.png" width="520" alt="A checkout page with a red banner: Payment failed: your card was declined.">
57
 
@@ -69,13 +68,13 @@ questions = {
69
  # next {"retry": 0.983, "wait": 0.004, "nothing": 0.014}
70
  ```
71
 
72
- v0.1 is a research preview. The weights are licensed under CC BY-NC 4.0.
73
 
74
  ## Use it
75
 
76
  ```bash
77
  pip install git+https://github.com/Xiaooolong/vev typesafe-sdk
78
- vev serve --model CountingSheep/vev-4b-lora # add --revision v0.1.0 to pin this release
79
  ```
80
 
81
  ```python
@@ -96,7 +95,7 @@ print(resp.answers["urgent"].noul)
96
 
97
  Accuracy on human-labelled data, by the kind of question:
98
 
99
- | Kind of question | Example | vev-4b-lora 0.1.0 |
100
  |---|---|---|
101
  | Jev-style decisions on text | JevBench public subset | 0.766 |
102
  | The same, on sources the model has not seen | kev transfer-v4 | 0.776 |
@@ -111,8 +110,8 @@ Accuracy on human-labelled data, by the kind of question:
111
  The text rows are public benchmarks; the image rows are judgment sets converted from labelled public data. Jev is
112
  more accurate on nimble (by 22 points), JevBench (10) and kev transfer-v4 (8); they tie on judgekit. Jev's API
113
  takes text only, so the image rows have no Jev comparison. Check Vev on a labelled sample of your own questions
114
- before relying on it. Significance tests and the comparison with the base model are in the [GitHub
115
- README](https://github.com/Xiaooolong/vev#detailed-results).
116
 
117
  ## Limitations
118
 
@@ -166,7 +165,6 @@ All Vev models and data:
166
  title = {Vev: Jev-like decision models that can also see images},
167
  author = {Wang, Xiaolong},
168
  year = {2026},
169
- url = {https://github.com/Xiaooolong/vev},
170
- version = {0.2.0}
171
  }
172
  ```
 
43
 
44
  <img src="https://raw.githubusercontent.com/Xiaooolong/vev/main/docs/doom.webp" alt="vev-4b playing Doom in real time: the game on the left, the eight slice probabilities and the chosen action on the right.">
45
 
46
+ vev-4b playing Doom in real time through a small harness: 8 yes/no questions per look, one request. Try vev-4b-v2
47
+ in the browser: [CountingSheep/vev](https://huggingface.co/spaces/CountingSheep/vev).
48
 
49
  This repository holds the LoRA adapter of [CountingSheep/vev-4b](https://huggingface.co/CountingSheep/vev-4b) (130
50
  MB). `vev serve` applies it to [Qwen/Qwen3.5-4B](https://huggingface.co/Qwen/Qwen3.5-4B) at load time, so it is the
51
+ smaller download if you already have the base model. The second generation,
52
+ [CountingSheep/vev-4b-v2](https://huggingface.co/CountingSheep/vev-4b-v2), has no adapter repository: its last
53
+ training stage trains a second adapter on merged weights, and only the merged result is published.
 
54
 
55
  <img src="https://raw.githubusercontent.com/Xiaooolong/vev/main/docs/checkout.png" width="520" alt="A checkout page with a red banner: Payment failed: your card was declined.">
56
 
 
68
  # next {"retry": 0.983, "wait": 0.004, "nothing": 0.014}
69
  ```
70
 
71
+ This is a research preview. The weights are licensed under CC BY-NC 4.0.
72
 
73
  ## Use it
74
 
75
  ```bash
76
  pip install git+https://github.com/Xiaooolong/vev typesafe-sdk
77
+ vev serve --model CountingSheep/vev-4b-lora
78
  ```
79
 
80
  ```python
 
95
 
96
  Accuracy on human-labelled data, by the kind of question:
97
 
98
+ | Kind of question | Example | vev-4b-lora |
99
  |---|---|---|
100
  | Jev-style decisions on text | JevBench public subset | 0.766 |
101
  | The same, on sources the model has not seen | kev transfer-v4 | 0.776 |
 
110
  The text rows are public benchmarks; the image rows are judgment sets converted from labelled public data. Jev is
111
  more accurate on nimble (by 22 points), JevBench (10) and kev transfer-v4 (8); they tie on judgekit. Jev's API
112
  takes text only, so the image rows have no Jev comparison. Check Vev on a labelled sample of your own questions
113
+ before relying on it. Per-set results are in `results/vev-4b/` of the [GitHub
114
+ repository](https://github.com/Xiaooolong/vev).
115
 
116
  ## Limitations
117
 
 
165
  title = {Vev: Jev-like decision models that can also see images},
166
  author = {Wang, Xiaolong},
167
  year = {2026},
168
+ url = {https://github.com/Xiaooolong/vev}
 
169
  }
170
  ```