--- license: apache-2.0 task_categories: - video-text-to-text - question-answering language: - en tags: - video - multi-hop-reasoning - rlvr - verifiable-rewards - reinforcement-learning size_categories: - 1K then A else B` — and the numbers are aggregated with `+ - *` into a single verifiable integer. Because *which* branch fires is never stated, the question cannot be answered without watching the video. --- ## ⚠ Read this before using the data **1. The corpus is a MIXED caption oracle.** 144 of 1,818 videos (7.9%) were captioned by `Qwen3.5-122B-A10B`; the other 1,674 by `Qwen3.5-397B-A17B-FP8`. This was forced, not chosen: the 397B has 32 attention heads, so vLLM accepts only TP ∈ {2,4,8}, and at TP=4 its 406 GB of weights need 101.5 GB/GPU — over an 80 GB card. One node in the cluster had 7 healthy GPUs and could not host it in any configuration. **The captioner is stamped per video in `_caption_meta.model`** (see `captions/captions_1080p.jsonl`), so the subset is identifiable. **Filter or audit it before any cross-video comparison.** **2. Generation never sees a frame.** The caption is the pipeline's *only* ground truth. A caption that fabricates poisons every question built on it. That is why resolution mattered so much (below) — but it also means **residual caption errors are the dominant remaining source of label noise.** **3. The graders were sampled (temperature 0.7), not greedy.** The keep/drop verdict is reproducible (per-request seeds) but not deterministic-by-design. --- ## The headline finding: 360p was the ceiling, not the captioner Earlier work on this corpus concluded that a 397B captioner *fabricated more on-screen text* than a smaller 122B. **That conclusion was an artifact of resolution.** At 640×360, distant signs, watermarks and brand text are physically unresolvable — every model was guessing at a blur. Clean A/B — **same model, same prompt, same segment boundaries; resolution the only variable** — on 9 concrete on-screen-text probes (word-boundary matched): | Qwen3.5-397B-A17B-FP8 | reads correctly | **misreads** | stays silent | |---|---|---|---| | @ 360p | 1 | **6** | 2 | | @ **1080p** | **4** | **1** | 4 | Fixed by resolution alone: `Southern Miss` → **Southern Utah**; `Cornstock` → **Comstock**; `SLOW AREA` → **REST AREA**. A fabricated `MIAMI` and a phantom HUD number `0397` both **vanish** — at 1080p the model goes **silent** rather than inventing, which is the failure mode that actually matters. --- ## The metric this dataset optimises: `blind_solver_rate` `blind_solver_rate` = accuracy of a solver that **never watches the video** and always takes the same branch. It is the number to keep *low*. | version | `then_rate` (healthy 0.35–0.65) | **`blind_solver_rate`** | keepers/video | |---|---|---|---| | v3.1 (360p) | 0.939 | **0.839** | 3.47 | | v3.2 (360p) | 0.627 | **0.261** | 1.76 | | **v3.4 (this release, 1080p)** | **0.507** | **0.133** | **4.47** | `then_rate` 0.507 is a near-perfect 50/50 split (all-THEN 0.133 vs all-ELSE 0.126) → **no branch bias**. **Answer diversity:** 964 distinct answers; the modal answer covers only ~1.3% of rows, so a constant-guess baseline scores ~1.3%. The 13.3% blind-solver figure comes from *branch strategy*, not from answers being guessable. ### ⚠ Do NOT credit resolution for the yield gains Judge survival rose 30.8% → 59.6% between v3.2 and v3.4, but **six things changed** (caption source, generator 122B→397B, 6→8 candidates/video, greedy→sampled decoding, a batched-rewrite refactor, corpus size). Per-candidate drop rates: | drop code | v3.2 (360p) | v3.4 (1080p) | Δ | mostly attributable to | |---|---|---|---|---| | `wrong_branch` | 52.7% | 23.4% | −29.4pp | generator strength + fix-loop | | `answer_mismatch` | 54.6% | 25.4% | −29.2pp | generator strength + fix-loop | | `not_stated` | 8.8% | 2.7% | −6.1pp | **caption fidelity** | | `unfaithful_quote` | 18.0% | 15.0% | −3.1pp | **caption fidelity** | | `attachment_error` | 0.3% | 1.5% | **+1.2pp** | ← counter-signal | | **caption-fidelity codes combined** | 27.1% | **19.1%** | **−8.0pp** | **resolution** | The headline −29pp is **generator strength, not resolution**. Resolution's isolated contribution is the **−8.0pp** on caption-fidelity codes. `attachment_error` *rose*: richer captions describe more objects, so there are more chances to mis-attach an attribute. --- ## Contents | path | what | |---|---| | `data/train-*.parquet` | **7,900 keepers.** verl RLVR schema: `prompt`, `videos`, `reward_model` (`ground_truth`), `extra_info`, plus flattened `answer` and `video_id`. | | `captions/captions_1080p.jsonl` | 1,818 videos × segment-level captions — **the oracle**. Carries `_caption_meta` (model, max_pixels, fps, max_frames), so the mixed-oracle subset is auditable. | **Videos are not mirrored here.** They are already public at [`Choiszt/video_res`](https://huggingface.co/datasets/Choiszt/video_res) under `videos_1080p/`. Join on `video_id`. ## Integrity (verified, 300-video sample / 1,331 rows) - ground truth empty: **0** - ground truth is a plain integer: **100%** — the verifiable-answer invariant - rows carrying a video reference: **100%** ## How it was built `caption → gen → judge → too-easy`, on 8× DGX (H100-80GB). | stage | model | detail | |---|---|---| | caption | `Qwen3.5-397B-A17B-FP8` (TP=8) + `Qwen3.5-122B-A10B` (TP=4, 144 videos) | native 1080p (`max_pixels=2073600`, nothing downscaled), `fps=1`, ≤24 frames/segment | | gen + rewrite | `Qwen3.5-397B-A17B-FP8` | 8 candidates/video; rewrite assigns per-hop polarity, de-circularises scene refs, LLM fact-checks each condition against the caption | | judge | `Qwen3.6-27B` | the only semantic gate — arithmetic is evaluated **in code**, never by the model, and the judge is never shown the target answer | | too-easy | `Qwen3.5-9B` | solves each question 8× on the full video; drops it if solved every time | **Funnel:** 1,818 videos → 13,938 candidates → 8,313 judge survivors (59.6%) → **7,900 keepers**. 51 videos (2.8%) yielded no candidates and are recorded, not silently dropped. Sampling (Qwen recipe): `temperature=0.7, top_p=0.8, top_k=20, min_p=0.0, repetition_penalty=1.0`. `presence_penalty` is **split by output type** — `1.5` on prose captions (it fixes runaway repetition), `0.0` on every JSON stage (there it penalises re-emitting the very keys each hop requires, and malformed JSON silently costs candidates). ## Known gaps - 53 of the original 1,871 videos have no 1080p source → excluded (a 360p subset would reintroduce a mixed-**resolution** oracle). - 51 videos yielded 0 candidates (2.8%) — ordinary generation/parse misses, *not* context overflow (verified: their caption-length distribution matches the successes). ## Citation Derived from the video corpus of [`Choiszt/video_res`](https://huggingface.co/datasets/Choiszt/video_res).