jbarney commited on
Commit
7f6fb7b
·
verified ·
1 Parent(s): 633d921

v1.2: options encoded side by side; answers do not depend on option order

Browse files
Files changed (4) hide show
  1. README.md +3 -0
  2. adapter/adapter_model.safetensors +1 -1
  3. config.json +21 -7
  4. head.pt +1 -1
README.md CHANGED
@@ -18,6 +18,8 @@ tags:
18
 
19
  # circuit-vl-4b
20
 
 
 
21
  **v1.1.** Same base, head and recipe as v1.0, with real photographs (Open Images V7, human-verified labels) added to the training data. "Vision grid v2" below names the dataset, not the model. See [Versions](#versions).
22
 
23
  A **System One** decision model for images: a state that carries one or
@@ -84,6 +86,7 @@ threshold.
84
 
85
  | tag | date | what changed |
86
  |---|---|---|
 
87
  | `v1.1` | 2026-09 | Trains on vision grid v2: the rendered cells plus three cells of real photographs. |
88
  | `v1.0` | 2026-09 | First release, rendered or scripted data only. |
89
 
 
18
 
19
  # circuit-vl-4b
20
 
21
+ **v1.2** (2026-09-22). The options of a choice question are encoded side by side, so the answer cannot depend on the order they are listed in: 0.0% of answers change under reordering on its grid (v1.1: 3.6%); grid .967 / ECE .021, POPE .907 / .069. `config.json` carries `"parallel_options": true`; serve it with the circuit repo's scorer, which applies the mask. See [Versions](#versions).
22
+
23
  **v1.1.** Same base, head and recipe as v1.0, with real photographs (Open Images V7, human-verified labels) added to the training data. "Vision grid v2" below names the dataset, not the model. See [Versions](#versions).
24
 
25
  A **System One** decision model for images: a state that carries one or
 
86
 
87
  | tag | date | what changed |
88
  |---|---|---|
89
+ | `v1.2` | 2026-09-22 | Options encoded side by side: order-stable answers. Same data and recipe as v1.1 otherwise. |
90
  | `v1.1` | 2026-09 | Trains on vision grid v2: the rendered cells plus three cells of real photographs. |
91
  | `v1.0` | 2026-09 | First release, rendered or scripted data only. |
92
 
adapter/adapter_model.safetensors CHANGED
@@ -1,3 +1,3 @@
1
  version https://git-lfs.github.com/spec/v1
2
- oid sha256:b332a595f1011483b1146bba91e96a7e72dd913d9afa723f79cf5807e0c50677
3
  size 132195448
 
1
  version https://git-lfs.github.com/spec/v1
2
+ oid sha256:9f221f558ffb8b2116bc038632b5775bd625c88de50476152d07957b32ea7a89
3
  size 132195448
config.json CHANGED
@@ -12,30 +12,44 @@
12
  "<|box_end|>",
13
  "<|fim_middle|>"
14
  ],
 
 
 
 
 
 
15
  "best": {
16
- "ece": 0.03463481445061533,
17
  "accuracy": 0.9671052631578947,
18
- "kl": 0.08838354297811024,
 
 
 
 
 
19
  "step": 600
20
  },
21
  "args": {
22
  "model": "Qwen/Qwen3-VL-4B-Instruct",
23
  "data": "data/vision/grid/train.jsonl",
24
- "out": "runs/circuit-vl-4b-v2",
25
  "epochs": 2,
26
- "batch": 2,
27
  "lr": 0.0001,
28
  "head_lr": 0.001,
29
  "rank": 16,
30
  "max_length": 1536,
31
  "val_frac": 0.1,
32
- "eval_every": 150,
33
  "limit": null,
34
  "seed": 0,
35
  "dtype": "bfloat16",
36
- "grad_checkpoint": false,
 
 
 
37
  "wandb": "s1proto",
38
- "run_name": "circuit-vl-4b-v2",
39
  "resume": false,
40
  "head": "pointer",
41
  "load_4bit": false,
 
12
  "<|box_end|>",
13
  "<|fim_middle|>"
14
  ],
15
+ "parallel_options": true,
16
+ "temperatures": {
17
+ "noul": 1.75,
18
+ "choice": 1.75,
19
+ "score": 1.0
20
+ },
21
  "best": {
22
+ "ece": 0.03482257006199742,
23
  "accuracy": 0.9671052631578947,
24
+ "kl": 0.09835513901895927,
25
+ "ece_by_type": {
26
+ "noul": 0.04722936921998076,
27
+ "choice": 0.031651588013538984
28
+ },
29
+ "ece_worst": 0.04722936921998076,
30
  "step": 600
31
  },
32
  "args": {
33
  "model": "Qwen/Qwen3-VL-4B-Instruct",
34
  "data": "data/vision/grid/train.jsonl",
35
+ "out": "runs/circuit-vl-4b-par",
36
  "epochs": 2,
37
+ "batch": 4,
38
  "lr": 0.0001,
39
  "head_lr": 0.001,
40
  "rank": 16,
41
  "max_length": 1536,
42
  "val_frac": 0.1,
43
+ "eval_every": 100,
44
  "limit": null,
45
  "seed": 0,
46
  "dtype": "bfloat16",
47
+ "acc_floor": 0.01,
48
+ "parallel_options": true,
49
+ "micro": 0,
50
+ "grad_checkpoint": true,
51
  "wandb": "s1proto",
52
+ "run_name": "circuit-vl-4b-par",
53
  "resume": false,
54
  "head": "pointer",
55
  "load_4bit": false,
head.pt CHANGED
@@ -1,3 +1,3 @@
1
  version https://git-lfs.github.com/spec/v1
2
- oid sha256:89e22c59ffe7cdc5c4b6b97ea424072237d1e68266b689f2db61ba7dd1b7fac6
3
  size 5244685
 
1
  version https://git-lfs.github.com/spec/v1
2
+ oid sha256:fb3e5b652c3a044b9156e995f75fcf7d09a6d87e26f07ac193b162784a8ecf87
3
  size 5244685