How do validators verify video clip assets without downloading entire datasets?

Validators sample rows from each miner interval (minimum 2 rows, maximum 35, targeting 25% of the interval) and download only the sampled clip and frame assets. They verify SHA256 hashes against declared values in the parquet rows and check sampled clip resolution is exactly 1280x720. Hard checks (schema, YouTube source verification, overlap policy) run on the full record set without downloading assets.

What happens if two miners submit overlapping video clips in the same interval?

Validators load a shared overlap index maintained by the owner-validator and prune any rows already seen in that index (row-level rejection, not full-interval rejection). For same-interval cross-miner conflicts on the same source, the miner with the earliest manifest creation timestamp wins; ties are broken lexicographically by hotkey.

Can validators check whether captions accurately describe video content?

Yes. Validators can optionally enable semantic caption validation, which samples rows and uses OpenAI (gpt-4o) or Gemini (fallback) to verify that captions match sampled multi-frame visual context. This check is disabled by default and requires configuring an LLM API key and enabling NEXIS_VALIDATOR_SEMANTIC_CHECK_ENABLED.

How often do miners need to upload new dataset packages?

Miners build and upload one package per 50-block interval. Every 50 blocks, the miner creates a new dataset.parquet and manifest.json keyed by the interval start block number and uploads to Cloudflare R2. Intervals are deterministic based on chain block height.

What does the owner-validator do differently from regular validators?

Owner-validators (identified by a specific hotkey) additionally publish accepted miner interval metadata bundles (dataset.parquet and manifest.json) to a shared bucket and maintain a global overlap index snapshot in JSON format. All validators read this shared index to prune duplicate rows. Owners can also run an independent worker to copy full assets to a separate owner bucket.

How are invalid miners prevented from earning rewards?

Validators fetch invalid hotkeys from an API (covering intervals in a rolling 500-interval window) and zero those hotkeys before submitting set_weights to the chain every 250 blocks. Hard failures (non-YouTube sources, schema errors, asset hash mismatches, source authenticity failures) also cause immediate interval rejection and contributor scoring penalties.

Subnet 70 (SN70) — Video Dataset Validation — Nexisgen

Building large-scale video datasets requires collecting clips from multiple sources, generating reliable metadata and captions, and verifying quality at scale. Manual validation is expensive and slow; centralized platforms create bottlenecks and single points of failure. A decentralized network of independent contributors—miners who collect and package clips, and validators who verify submissions—can perform this work in parallel while maintaining quality through cryptographic incentives and on-chain scoring.

Nexisgen implements this as a Bittensor subnet. Miners collect YouTube video clips, generate captions from timeline-sampled frames, and upload interval-based dataset packages (parquet files with manifests) to decentralized storage every 50 blocks. Validators download submissions, perform multi-layer checks including schema validation, asset integrity verification, and optional semantic/category checks, then score miners and submit weights to the chain. Invalid miners are zeroed before weight submission.

The subnet enforces dataset quality through hard checks (YouTube source verification, no overlapping clips from the same source within 4.5 seconds, caption lexical standards, SHA256 asset verification) and optional semantic validation. An owner-validator maintains a shared overlap index so all validators can prune duplicate rows across miners without rejecting entire intervals. Same-interval cross-miner conflicts are arbitrated by earliest manifest creation timestamp.

Machine learning teams can build video training datasets without managing clip collection infrastructure, leveraging miners' diverse sourcing while validators ensure no duplicates or low-quality clips enter the dataset.
Computer vision researchers can validate caption-to-visual alignment at scale without hand-labeling, using optional semantic checks that verify captions match multi-frame context from randomly sampled clips.
Data engineers can enforce category-specific dataset constraints (nature/landscape/scenery) through validator category checks, which require manifest metadata and perform strict vision evaluation on middle timeline frames.
Organizations needing video datasets with clear provenance can audit clip sources and creation timestamps on-chain, with source URLs and proof metadata embedded in every row and cross-miner overlaps automatically pruned.
Node operators can earn rewards by running validators that verify dataset integrity without storing full assets, using sampled asset verification and cryptographic scoring to measure contribution fairly.

Nexisgen is a Bittensor subnet where miners produce interval-based video clip datasets (with captions and metadata) and validators verify dataset integrity, score miners, and submit chain weights. Miners collect YouTube video clips, generate captions from timeline-sampled frames, and upload parquet datasets with manifests to Cloudflare R2 storage every 50 blocks. Validators download submissions, perform schema/integrity/asset verification, run optional semantic and category checks, prune overlaps using a shared index, and compute miner scores based on validation sampling.

Miners run a continuous loop that: (1) discovers new 50-block intervals from chain, (2) collects YouTube video clips using yt-dlp, (3) generates clip records with captions from timeline-sampled frames, (4) writes interval package (dataset.parquet + manifest.json), (5) uploads to Cloudflare R2 bucket. Validators run a continuous loop that: (1) discover active miners via metagraph and committed on-chain credentials, (2) download each miner's interval packages, (3) validate via multi-layer checks (manifest identity, integrity, schema, asset verification, sampled semantic/category validation), (4) prune rows using shared overlap index from owner validator, (5) arbitrate same-interval cross-miner conflicts by earliest manifest timestamp, (6) compute per-miner scores, (7) accumulate scores over 250-block windows and submit set_weights to chain. Owner-validator is a special validator instance that additionally publishes accepted metadata bundles and overlap index snapshots to shared buckets for other validators to use.

Machine learning teams can build video training datasets without managing clip collection infrastructure, leveraging miners' diverse sourcing while validators ensure no duplicates or low-quality clips enter the dataset.
Computer vision researchers can validate caption-to-visual alignment at scale without hand-labeling, using optional semantic checks that verify captions match multi-frame context from randomly sampled clips.
Data engineers can enforce category-specific dataset constraints (nature/landscape/scenery) through validator category checks, which require manifest metadata and perform strict vision evaluation on middle timeline frames.
Organizations needing video datasets with clear provenance can audit clip sources and creation timestamps on-chain, with source URLs and proof metadata embedded in every row and cross-miner overlaps automatically pruned.
Node operators can earn rewards by running validators that verify dataset integrity without storing full assets, using sampled asset verification and cryptographic scoring to measure contribution fairly.

Validators sample rows from each miner submission (sample_rate=0.25, min=2, max=35 rows per interval; for each miner, if row_count<=10 sample all, else sample 20% capped at 10). Hard checks run on full dataset (schema, source URLs, overlap, captions). Sampled rows undergo additional expensive checks (asset verification, semantic caption validation, category validation).

Accepted intervals contribute to miner score; rejected intervals contribute zero. Failure types include non-YouTube sources, schema violations, asset hash mismatches, caption mismatches, overlap violations, source authenticity failures, and category validation failures. Validators maintain rolling 10-interval failure lookback and can dampen/zero recent hard-failure miners.

API-backed invalid hotkey lists (interval_id-500 to interval_id) are fetched and applied before weight submission.

Subnet 70

Registration Activity

Subnet 70

Subnet Configuration

Market & Economics

For Miners

For Validators & Stakers

Registration Activity

What does Subnet 70 do?

Use Cases

About

Subnet History

Frequently Asked Questions

Development Activity

Recent Activity

Hardware Requirements

What does Subnet 70 do?

Use Cases

About

Subnet History

Frequently Asked Questions

Subnet Configuration

Market & Economics

For Miners

For Validators & Stakers

Hardware Requirements

Miner Setup

Validator Setup

Overview

Registration Activity

Participate

Overview

Subnet Configuration

Market & Economics

For Miners

For Validators & Stakers

Registration Activity

What does Subnet 70 do?

Use Cases

About

Subnet History

Frequently Asked Questions

Participate

Development Activity

Recent Activity

Hardware Requirements

What does Subnet 70 do?

Use Cases

About

Subnet History

Frequently Asked Questions

Subnet Configuration

Market & Economics

For Miners

For Validators & Stakers

Hardware Requirements

Miner Setup

Validator Setup