Data Feeds · For AI Research Labs

Your model has read
a billion lines of code.
It has never seen one fail.

Code corpora are scraped from repositories at rest — the happy path, already merged, already green. Nothing in them shows a container OOM at 3am, a schema drift two releases later, or an agent quietly exfiltrating to an egress it was never supposed to reach. We run the tools. We record what actually happens.

🟢 Positive samples
Clean executions and the exact outputs they returned
🔴 Negative samples
Timeouts, schema violations, OOM kills — the rejected half of every DPO pair
⚠️ Security events
Runtime egress attempts and OWASP LLM findings, caught in the sandbox
Drawn from 55,108 sandbox runs across 1,812 tools · corpus grows every swarm cycle
One row of it

This is what a negative sample looks like.

Not a label someone wrote. A container that was built, booted, probed, and died — with the input that killed it and the trace it died on, captured at the moment of failure.

// one line from the DPO feed — rejected branch
"tool": "agentmemory", "commit_sha": "a3f9c21e0b47",
"input": "{ query: 'recall session 7', top_k: 512 }",
"outcome": "failure", "error_code": "health_check_timeout",
"peak_memory_mb": 2047.8, "init_latency_ms": 31402,
"owasp_flags": ["LLM10"], "egress_attempts": 3,
"pass_k_1": 0.0, "sample_type": "machine_verified"
Every field above is emitted by the sandbox itself — sample_type: machine_verified means no human, and no vendor, touched it.

That row isn't an orphan. It belongs to a life.

Every JSONL line above is one heartbeat of a repository the swarm has been following for months — through releases, relicenses, CVEs, cost, verdicts and revocations. Name any asset in the registry and pull its whole thread, live from the same D1 tables the paid feeds export.

Assembling eight histories…

Granular I/O traces

The exact synthetic input sent to the agent and the exact JSON/HTTP output returned, sequence by sequence, credential-redacted at capture.

Time-series regression

Teach models how software rots. Memory bloat and CVE introductions tracked across commit SHAs and release tags on a 7-year immutable timeline.

Hugging Face native

Pre-formatted commercial-use JSONL. Stream it straight into your pipeline with the datasets library — no reshaping.

Nobody in this dataset was asked how they were doing.

Benchmark corpora get gamed because they are self-reported: a vendor runs their own tool, on their own harness, and publishes the number they like. Every row here comes from an adversarial third party — our swarm cloned the repo, built the image, ran it in a network-isolated container it did not control, and wrote down what happened. A tool cannot flatter itself into this dataset, and it cannot opt out of the run that made it look bad.

Access tiers

Priced for the value of the corpus.

Pro
$249/mo
  • Weekly SFT/DPO dataset updates
  • Trailing 30-day behavioral corpus
  • sft / dpo / security JSONL feeds
  • Commercial-use data license
Labs
Enterprise
Custom/mo
  • Real-time Kafka / S3 streaming
  • Full immutable 7-year execution archive
  • Custom threat taxonomies
  • Priority pipeline + SLA

Delivered under the VerditNxtGen Commercial Data License. Sample and Pro feeds are token-gated to your API key; Enterprise streaming is provisioned per contract.