xvreason/vbvr-thinkgen-1m

VBVR ThinkGen 1M

The v8 ThinkGen prompts for all 1,000,000 training videos of VBVR (100 generator families × 10,000 clips). It is the full-corpus counterpart of xvreason/vbvr-thinkgen, which carries the same revision for a balanced 99,958-clip subset (1,000 clips per family).

This repo holds prompts only. The videos are the unmodified clips of Video-Reason/VBVR-Dataset; every video path in metadata.csv resolves inside that repo's extracted tars/<family>.tar.

How the prompts were made

The prompts were not regenerated. The revision that produced the 100k snapshot is a chain of deterministic, per-family text rules, and the same rules were replayed over all 10,000 clips of every family:

Stage	Input	Rules	Families	Prompts changed (1M)
base	original VBVR	May-apple/VBVR-Reorganized new_prompts_all.csv	29	—
v5	base	ThinkGen v3–v5 additive de-leak (phrase rules, one whole-prompt rewrite)	26	260,000
v6	v5	owner-review pass, 2026-09-08	30	300,000
v7	v6	reasoning-prompt checklist desk pass, 2026-09-10	21	204,154
v8	v7	ground-truth pass against the generator parameters, 2026-10-06	12	86,454

The xvreason/vbvr-thinkgen card describes what each stage changes and why.

In total, 674,071 of the 1,000,000 prompts (69 of 100 families) differ from the original VBVR text, and 650,225 differ from VBVR-Reorganized.

Two parts of the chain depend on individual clips and were rebuilt for the full corpus rather than copied:

• Per-clip corrections. G-19's "horizontal row" correction (4,927 clips whose ground truth is a row) and O-3's two-row wording (5,013 clips with symbols laid out in two rows) are selected from each clip's generator parameters in VBVR's data/metadata.parquet. The lists are in lists/.
• G-18's colour removal was a token-slot edit on the 100k's dominant template. Here it is applied as an equivalent phrase rule so that it also reaches templates that occur only in the full corpus.

Verification

• The replay reproduces the published subset byte for byte. Restricted to the 99,958 clips of the balanced subset, every stage's output has the SHA-256 of the corresponding published snapshot: v5 a8f156b0…, v6 9049cdac…, v7 11a7a6b8… and v8 681ab1c3…. Those clips are listed in balanced_100k_videos.txt.
• No family was refused on the other 900k clips. Each rule file contains must-match, must-be-present and must-be-absent checks plus a mechanical defect scan (articles, double spaces, fragments and similar). If a family fails any check, all of its rows are kept at the previous stage.
• The other 900k clips behave like the subset. Per family and per stage, the fraction of prompts changed in the 900k clips is within 0.03 of the fraction in the subset. The defect rate does not increase, and the longest prompt is the same, 666 characters.
• Prompt variants new at 1M (10,801 clips across six families) are new value combinations inside sentence frames the subset already has. Examples are other shape and colour words, or three and four traffic lights instead of two. None of them introduces a new sentence structure.

The blind-solvability and sighted A/B measurements reported for the subset were not re-run on the extra clips. Because the rules produce the same sentence frames, those measurements apply to the full corpus.

dataset_manifest.json records the hashes and counts, and build_report.json records the stage-by-stage, per-family outcome.

Full-corpus audit

After the build, every family was checked on all 10,000 of its clips. Each statement a prompt makes (counts, colours, positions, rules, tie-breaks, stop conditions) was tested by code against that clip's generator parameters, with the balanced subset and the other 900k clips counted separately.

• No text defect or reasoning-gate failure occurs only outside the subset. Every issue found in the other 900k clips also occurs in the subset, at a similar rate.
• G-18 and G-12's route convention was checked on the videos. Their prompts say the agent moves "horizontally first and then vertically". The parameters do not record the route, so the agent was tracked in every ground-truth video of both families (code/scripts/probe/route_check_1m.py). The route was recovered in 20,000 of 20,000 clips, and every route is a shortest path. Every leg that has more than one shortest route goes horizontally first: 8,257 of 8,257 legs in G-18 and 31,548 of 31,548 in G-12. Per-clip results are in route_check/.

Clips whose ground truth contradicts the prompt

flagged_clips.csv lists 1,751 clips (0.18% of the corpus) where the ground-truth video contradicts an explicit statement in the prompt. These rows are kept in metadata.csv, so filter them out if you want a contradiction-free set. All of these issues come from the original VBVR generators. Most of them also occur in the balanced subset (in_balanced_100k).

Family	Issue	Clips (subset / rest)
G-12	the horizontal-first route steps on the end point before the last reward, although the prompt says to collect every reward before reaching it	98 / 854
O-47	the prompt says "Complete in exactly N moves … Do not make extra moves", but fewer moves solve the puzzle	47 / 513
G-16	a straight leg runs over the end square before the last block is visited	10 / 115
G-146	a circled "square" has unequal sides, or an equal-sided shape is not circled	4 / 45
O-31	the black ball eats a ball that is not smaller than itself	5 / 37
G-16	the recorded block order is not the shortest	0 / 8
G-30	the highlighted element is not the requested maximum or minimum	2 / 5
G-198	a marked triangle has zero area	0 / 5
G-31	the video takes more steps than the minimum	0 / 3

The list does not cover clips with more than one valid answer (for example, several equally short routes) or clips where the final frame equals the first. The prompt is still true of the video in those cases.

Files

File	Contents
metadata.csv	video,prompt: 1,000,000 rows in DiffSynth-Studio metadata format
balanced_100k_videos.txt	the 99,958 clips of the balanced subset (xvreason/vbvr-thinkgen)
lists/*.json	per-clip selections used by the v8 G-19 and O-3 rules
flagged_clips.csv	video,family,issue,description,in_balanced_100k: clips whose ground truth contradicts the prompt
route_check/*.json	per-clip route tracked in every G-18 and G-12 ground-truth video
code/	the build script, the v6–v8 rule files and the route checker
dataset_manifest.json	hashes, counts and stage gates
build_report.json	per-stage, per-family build result

Prompts can contain embedded newlines. Read metadata.csv with an RFC-compliant CSV parser, not with line-oriented shell tools.

Using it with the videos

hf download xvreason/vbvr-thinkgen-1m --repo-type dataset --local-dir vbvr-1m
hf download Video-Reason/VBVR-Dataset --repo-type dataset --include "tars/*" --local-dir vbvr-src
cd vbvr-1m && for t in ../vbvr-src/tars/*.tar; do tar -xf "$t"; done

The video values in metadata.csv then resolve from vbvr-1m/. The source tars also contain first and final frame PNGs and metadata.json files, which training does not need. To keep only the videos, extract with --wildcards '*/ground_truth.mp4'.

Scan for decodability before training. The 100k subset leaves out 42 G-134 clips because our own interrupted extraction truncated them, and one such file crash-looped a training run. Run an ffprobe pass over whatever you extract.

Provenance and license

Videos and original prompts come from Video-Reason/VBVR-Dataset, and the base prompt revision comes from May-apple/VBVR-Reorganized. Both are Apache-2.0, and this derivative is released under the same license. The build script and rules are in code/. To rebuild metadata.csv, run code/scripts/data/build_vbvr_1m.py on VBVR-Reorganized's new_prompts_all.csv and VBVR's data/metadata.parquet. Pass --published-dir a folder holding the metadata.csv of xvreason/vbvr-thinkgen tags v5–v8, saved as v5.csv–v8.csv. The script checks its output against those files.