20 corpus collections across video, audio, image–text pairs, question banks, and code-agent tasks. Prepared for continued pre-training, multimodal alignment, SFT, RL, and evaluation. Scope is confirmed before delivery; sample packs and data cards are shared after review.
20 collections today, each shipped with a manifest, a data card, and a field schema.
Short-video, branded video, vlog, 4K stock and camera-movement footage, sports video with three-level temporal annotation, 4K game screenshots, egocentric sessions with seven annotation types, and video SFT clips — each item with structured metadata.
Nine subjects across primary, middle, and high school, plus 16 vocational exam categories. Every item carries stem, options, answer, worked solution, and knowledge point.
Repository-level tasks: one open-source issue per package, with a runnable sandbox image, reference patch, test patch, and an automatic grading script.
E-commerce image–text pairs and attribute-extraction SFT records with field-level evidence chains. The current release holds 2,808 records.
Music stems split by instrument — lead vocal, harmony, accompaniment — delivered as track-organized WAV sets, including a full 41-stem reference track.
Sort by modality, then by training stage: the same 20 collections, one catalogue, two views.
| Modality | Continued pre-training & alignment | SFT | RL & agents | Evaluation |
|---|---|---|---|---|
| Text · question banks | — | 9 subjects across three school stages and 16 vocational categories with worked solutions; attribute extraction with evidence chains | — | Held-out question sets |
| Image–text | E-commerce pairs with attributes and titles | Attribute extraction with field-level evidence chains | — | — |
| Image | — | — | — | 4K game screenshots for art-style checks |
| Video | Short-video, branded, vlog, 4K and camera-movement footage; egocentric sessions with seven annotation types | Video-understanding SFT clips | — | Held-out clips with reference labels |
| Audio | Instrument-split stems — vocal, harmony, accompaniment — as track-organized WAV | — | — | — |
| Code · agents | — | Issue-to-patch trajectories with reference patches | Sandboxed tasks with automatic grading | Automatic grading scripts |
Swipe sideways for every stage →
Five steps from scope confirmation to signed delivery.
Confirm collection, volume, schema, and delivery format.
Register raw files, IDs, versions, and integrity checks.
Parse, clean, annotate, and structure to the agreed schema.
Technical validation, content sampling, cross-field consistency.
Data, manifest, checksums, and acceptance records.
Scope, volume, format, and timeline are confirmed per project.
| Item | What ships |
|---|---|
manifest.json | Every file with name, size, date, and collection ID. |
| UTF-8 JSONL | One record per line for text and annotations. |
| Media files | Original-resolution video and audio, plus proxy files for review. |
| Packaged tasks | docker/ (OCI layout) and json/ (task, patches, grading script). |
| Documents | Data card, field schema, and a PDF spec for each collection. |
Parameters not yet fixed are confirmed against your project scope.
Three notes from the team, written for the people who own the model.
Why collection and schema decide quality before training starts.
How to weigh data quality against the cost of a failed run.
Turning service records into training material without a new collection.
Tell us the modality, training stage, and volume. We reply within one business day. Sample packs, data cards, and field schemas are shared after your request is reviewed.
Yes. Most projects start with one collection or one modality, then extend after acceptance.
After the scope conversation we share a sample pack, the data card, and the field schema.
Each collection is delivered as a structured dataset with a fixed schema. Format and volume are confirmed for your project.
Annotation schema, sampling, and delivery format can be adjusted to your pipeline.
Yes. We scope new collections with the same production process before quoting.