MiniMax_Distill_Data

5,021 English text-to-video prompts paired with a legal MiniMax-H3 frame count, spread evenly over all 14 durations the model supports. Pure text-to-video: prompts that require a reference image have been removed, and Seedance/Midjourney-specific markup has been stripped so the text is what MiniMax-H3's conditioner actually wants. 424 prompts are rewritten into MiniMax-H3's official T2VA prompt schema, 218 of them using the <d> dialogue tokens that drive the audio branch. Ready to feed straight into a MiniMax-H3 pipeline for distillation / cache-LoRA / step-caching training runs.

Derived from GokuScraper/seedance-2-prompts-datasets (CC BY 4.0) — 8,755 scraped Seedance 2.0 generations with bilingual prompts and the spec of the resulting video. Only the English prompt and the duration are carried over; no videos or images.

Fields

field	type	description
id	str	MMD_00000 …
prompt	str	English prompt, normalized (NFKC, full-width punctuation folded)
num_frames	int	Frame count to pass to MiniMax-H3, always of the form 17n + 5
duration	float	num_frames / 24, seconds
duration_source	str	how this row got its length — see below
prompt_format	str	h3_schema (official three-field layout) or plain prose
has_dialogue	bool	prompt contains <d>…</d> spoken content
timing_bound	bool	whether the original text was tied to one specific length
mobility	str	free / movable / pinned — how the row could be rebalanced
prompt_length	int	characters in prompt
source_id	str	Row id in the upstream dataset
source_duration	float	Duration of the upstream Seedance clip, seconds

duration_source values:

value	rows	meaning
captured	1,607	length comes from the source clip, prompt text untouched
reassigned	1,638	prompt names no length, so it was simply moved to another bucket
rescaled	1,053	timed shot list linearly rescaled to the new length
retimed	750	stated length rewritten (15-second → 9-second)

Duration grid

MiniMax-H3 generates 5–15 s at 24 fps and its video VAE only encodes frame counts of the form 17n + 5, so exactly 14 requests are legal. raw is what the source durations alone give:

frames	seconds	rows		frames	seconds	rows
124	5.167	361		243	10.125	361
141	5.875	360		260	10.833	360
158	6.583	356		277	11.542	358
175	7.292	359		294	12.250	359
192	8.000	358		311	12.958	357
209	8.708	361		328	13.667	360
226	9.417	358		345	14.375	353

Every bucket holds 353–361 rows (7.03–7.19%); the raw scrape had 69.5% of rows at 345 frames and 3 at 158. There is no 4-second setting — the shortest clip MiniMax-H3 can produce is 124 frames (5.167 s), so upstream clips below that were clamped up rather than dropped.

How it was balanced

Seedance's UI presets are 5/10/15 s, so the raw durations are tri-modal: 69.5% of rows landed on 345 frames and the 158-frame bucket had 5. Training a duration-conditioned model on that would mean running nearly every step at the longest — and slowest — setting.

Every prompt was classified by a locally-run Qwen3.5-9B into bound (the text is tied to one length) vs free, then each row was placed by mobility:

• free (no length mentioned) — moved verbatim.
• movable (bound, but the binding is mechanical) — the text is rewritten. A stated length becomes the new one (FORMAT: 15s / 140 BPM → FORMAT: 9s / 140 BPM, with a/an corrected), and a timed shot list is linearly rescaled (0-4 seconds: … 9-15 seconds: … at 15 s becomes 0-1.9 seconds: … 4.4-7.3 seconds: … at 7.3 s).
• pinned (421 rows) — the rewrite could not be verified, so the row keeps its captured length.

Rewrites are self-verifying: after rewriting, the prompt is re-scanned and rejected if any timing still refers to a moment past the new clip length. 0 of the 3,441 modified rows reference a time beyond their duration. Durations that belong to the fiction rather than the edit (a 60-second countdown clock, a 24-70 mm lens) are left alone.

A regex pre-filter settles the clear cases and the model is asked only about the ambiguous ones. Regex alone is inadequate in both directions: it misses FORMAT: 15s / 140 BPM and fires on incidental strings like a man in his late 20s, early 2000s VHS, 80s movie scene — all of which the model correctly called free.

Mean duration is 9.77 s (raw: 12.84 s); a full capture is ~13.7 video-hours.

Formatted for MiniMax-H3

MiniMax-H3's t2va text encoder takes "the prompt verbatim, with no chat template and no special tokens" and tokenizes with add_special_tokens=False, so every character becomes conditioning tokens — including anything the author addressed to Seedance's UI rather than to the model. That markup was removed:

removed	rows
HTML entities from the scrape (' → ')	85
UI/scrape chrome (### TEXT TO TRANSLATE ###, Prompt 👇)	26
Midjourney flags (--v 6.1, --stylize 250, --sref …)	22
--ar 9:16 → rewritten as , 9:16 aspect ratio	21
tool credits (Made in @runwayml, via @capcutapp)	14
hashtags	9
tool spec headers (`Seedance 2	7s cinematic AI video …`)

Ordinary filmmaking language is kept: camera moves, lens and film stock, aspect-ratio and resolution wording, shot lists with timings, style and director references. Aspect ratio is passed to the pipeline as height/width, but the words still describe the shot, so they stay.

Verified against the real MiniMax-H3 tokenizer: the only special token ids emitted are the intentional <d>/</d> pair, and lengths are 8–3,259 tokens (median 201) against a 262,144 limit — nothing needs truncating.

The official H3 prompt schema (prompt_format = "h3_schema")

MiniMax-H3's model card notes it "adds several special tokens, such as <d>, to the tokenizer", and its prompt writing guide specifies a three-field T2VA layout that H3-Context-IR normally produces. <d> and </d> are genuine single tokens (ids 151669 / 151670) that mark spoken content for the audio branch — dialogue written as ordinary quotes does not reach it.

The Seedance scrape uses none of this: 0 rows had any schema field, <d> tag or (S1) speaker id. So 424 prompts (~30 per duration) were rewritten by a locally-run Qwen3.5-9B into:

integrated_multimodal_description: [Shot 1] Live-action, cinematic, a medium shot frames a man in
brown clothes setting down a bamboo basket on snow. The middle-aged man with a calm, raspy voice
(S1) says: <d>[English] Here's a bunch of chives for you, hope you can survive the winter.</d>
[Shot 2] At 00:01.700, the camera cuts to a courtyard in warm natural light...

overall_soundscape: Snow crunches underfoot while a low wind moves through bare branches...

non_diegetic_music: Sparse guzheng notes at a slow tempo, fading out before the final shot.

Rewrites carry the guide's conventions: [Shot N] tags, strictly increasing At 00:03.500 cut times, the guide's camera-motion vocabulary (Push In, Truck Left, Arc Shot, "with small amplitude", "at slow speed"), stable (S1)/(S2) speaker ids with voice descriptions, and <d>[English] ...</d> for speech.

Every rewrite is structurally validated before it is accepted — field order, [Shot 1] opening, consecutive shot numbering, cut times inside the clip, balanced <d> tags each carrying a language tag, no invented dialogue for a prompt that had none, no tool names, no reference-image wording — and then re-checked against H3's own tokenizer to confirm <d> survives as its real token id. Acceptance was 86.5% (424/490); rejected rows keep their original prose. Verdicts for all 490 attempts, including failures and their reasons, are in schema_verdicts.jsonl.

	rows
prompt_format = "h3_schema"	424
of those with <d> dialogue	218
of those with (S1) speaker ids	202
prompt_format = "plain"	4,597

Because H3 reads the prompt verbatim, both formats are valid input; the schema rows simply exercise the conditioning the model was designed around.

A locally-run Qwen3.5-9B audited all 5,129 prompts as a backstop for markup the rules did not know about. It is a detector, not an authority — it echoed examples from its own instructions and flagged legitimate shot lists — so only findings that could be confirmed mechanically were acted on. It earned its keep by catching 78 prompts with CJK reference-image markers (@图片4, @画像1, リファレンス画像) that the English-only image filter had missed.

Text-to-video only

The Seedance scrape mixes text-to-video with image-to-video and reference-to-video posts, so many prompts point at an input this dataset does not have — @[image1], <<<image_1>>>, "the character in Image 1", "using the uploaded reference image", "animate the provided first frame". Training a text-to-video model on those teaches it to invent a subject it was never shown, so 1,219 rows (19.2%) were removed.

Unambiguous markers are matched by regex (1,197 rows). The 47 prompts whose only signal was first-frame language went to Qwen3.5-9B, because "start frame" can mean either an supplied image or just the opening shot of the generated video — it kept 25 ("the camera is already moving fast from the first frame") and dropped 22 ("animate the provided first frame"). image_verdicts.jsonl records the verdict for all 6,348 pre-filter rows, including the dropped ones.

Filtering

Starting from 8,755 upstream rows:

step	dropped
source duration outside 4.0–15.5 s (or garbage)	1,266
duplicate prompt (case/whitespace-insensitive)	976
prompt not actually English (CJK body text)	152
prompt too short (<40 chars) or too long (>12k)	13
requires a reference image / video	1,219
CJK reference-image markers (@图片4, @画像1)	80
nothing left after removing tool chatter	1
asset handles / tool meta-instructions	27
kept	5,021

Split: 4,742 train / 279 eval, ~20 eval prompts per frame-count bucket.

A second cleaning pass removed Seedance residue the first pass missed: [Seedance Parameters: …] blocks, --character_ref handles, upload ids like hf_20260513_135105_6944185e-…, and direct address to the old tool (Use Seedance 2.0 to generate …, Created with Seedance 2.0). Rows whose whole prompt is a prompt-engineering instruction ("You are an expert video prompt engineer specialized in Seedance 2.0…") are dropped rather than surgically edited. About 40 prompts still contain the word "Seedance" as fiction — "the opulent yet chilly Seedance headquarters", a parody short — and those are scene content, so they stay.

Usage

from datasets import load_dataset

ds = load_dataset("prodia/MiniMax_Distill_Data", split="train")
row = ds[0]
pipe(prompt=row["prompt"], num_frames=row["num_frames"], height=768, width=1344)

To train only on untouched, source-derived durations:

ds = ds.filter(lambda r: r["duration_source"] == "captured")

Reproducing

build_distill_data.py builds the unbalanced set from the upstream dump; filter_image_prompts.py writes image_verdicts.jsonl; normalize_for_h3.py and strip_seedance.py remove foreign markup; classify_prompt_timing.py (Qwen3.5-9B, shards over N GPUs) writes timing_verdicts.jsonl; verify_h3_prompts.py is the LLM audit; retime_prompts.py holds the duration-rewrite logic; rebalance_v2.py does the assignment; rewrite_h3_schema.py produces the schema rows and schema_verdicts.jsonl.

Caveats

• 3,441 rows have machine-edited prompt text. The edits are mechanical and verified for internal consistency, but they are not human-reviewed. duration_source == "captured" means the length is untouched; markup normalization may still have applied to those 166 rows.
• Plain-format prompts still contain their own structure ([Style], Shot 2 (2s–5s), Subject: headers). That is authored description, not foreign syntax, and MiniMax-H3 reads it as text.
• The 424 h3_schema rows are machine-authored from the original prompt. They follow the official guide and pass structural validation, but the wording is the 9B model's, not the original author's. Filter on prompt_format == "plain" to avoid them.
• <d> content is only as good as the dialogue in the source prompt; the rewriter is forbidden from inventing speech, so prompts without dialogue have no <d> tags.
• Some rows carry an inconsistency that was already present upstream (text referring past the clip's own length). These were left untouched rather than silently rewritten.
• timing_bound and the image filter are a 9B model's judgement plus regex, not ground truth. All verdicts are published so either decision can be redone differently.
• 25 kept prompts still say "first frame" or "start frame" in the sense of the opening shot of the generated video. They were individually cleared by the model, not missed.
• Prompt length is long-tailed: median ~930 characters, max ~12k. Truncate to your text encoder's limit.

License & attribution

CC BY 4.0, inherited from the upstream dataset. Attribute GokuScraper/seedance-2-prompts-datasets. Copyright in the original generated videos belongs to their creators; this repository contains prompt text and metadata only.