SelfThinker Data: Reliability-Verified Visual Thought Pools & Benchmark Suite

This repository contains the reliability-verified intrinsic visual-thought training sets and the benchmark TSV files for training and evaluating SelfThinker.

• Models: https://huggingface.co/jankin123/SelfThinker-7B
• Data: https://huggingface.co/jankin123/SelfThinker
• Code: https://github.com/zhangquanchen/SelfThinker

---

1. Reliability-Verified Training Sets

Synthesized multi-object visual-search problems (384x384 px canvases with 5–8 non-overlapping shapes across 5 shapes and 8 colors) where candidate intrinsic visual thoughts ${v^{(k)}}_{k=1}^{6}$ are scored by conditional pointwise mutual information:

$$u(v)=\log p_{\theta }(d\mid C(I,q,v))-\log p_{\theta }(d\mid C(I,q))u(v)\; =\; \backslash log\; p\_\backslash theta\backslash big(d\; \backslash mid\; C(I,\; q,\; v)\backslash big)\; -\; \backslash log\; p\_\backslash theta\backslash big(d\; \backslash mid\; C(I,\; q)\backslash big)$$

and retained when the top candidate $v^\star$ exceeds the item-specific 95th-percentile mismatched threshold $Q_{0.95}(s_i)$.

Directory	Verification Direction $d$	Verified Items	Parquet File	Command
audit_pass_l0_vsearch_trainonly/	Ground-truth answer $a^\star$	1,092 / 1,500	Visual_Search/data/train-00000-of-00001.parquet (111.7 MB)	scripts/train_selfthinker.sh answer
audit_pass_l1_vsearch_trainonly/	Label-free rollout majority vote $\tilde{a}$	1,074 / 1,500	Visual_Search/data/train-00000-of-00001.parquet (107.4 MB)	scripts/train_selfthinker.sh vote

---

2. Benchmark Evaluation Suite (lmudata/)

lmudata/ provides all self-contained TSV files (with embedded base64 images) used by the seven-benchmark evaluation script (scripts/eval_nine.sh) and the multi-layer image-condition probe (scripts/eval_image.sh):

Benchmark	TSV File(s)	Items	Size	Aggregation Rule
VSP	lmudata/VSP_maze_task_main_original.tsv	600	144.3 MB	Main maze task accuracy
ChartQA	lmudata/ChartQA_h_bar.tsv, lmudata/ChartQA_v_bar.tsv	444 + 382	35.7 MB, 19.7 MB	Mean of horizontal and vertical bar subsets
V*	lmudata/VStarBench.tsv	191	162.0 MB	Direct accuracy
MMVP	lmudata/MMVP.tsv	300	4.2 MB	Direct accuracy
SAT	lmudata/SAT_circular.tsv	150	354.1 MB	Circular evaluation split accuracy
BLINK	lmudata/BLINK.tsv	1,901	538.4 MB	Overall accuracy across 14 subtasks
CV-Bench	lmudata/CV-Bench-2D.tsv, lmudata/CV-Bench-3D.tsv	1,438 + 1,200	139.9 MB, 158.2 MB	Mean of 2D and 3D subsets

---

Download & Usage

git clone https://github.com/zhangquanchen/SelfThinker.git
cd SelfThinker

huggingface-cli download jankin123/SelfThinker --local-dir data
huggingface-cli download jankin123/SelfThinker-7B --local-dir models