Questions and answers
- What is Gerra?
Gerra builds the data and RL environments frontier AI models train on. It originates documents, codebases and business records from real work at 4,000 companies, and builds RL environments, evaluations and agent trajectories on top of them and from scratch, including synthetic spec-to-repo coding tasks. Gerra has processed 800,000 documents, holds over 6,000 codebases with about 1.4 million commits, and has delivered 64,000 graded back-office tasks and 96,000 computer-use trajectories.
- Where does Gerra's data come from?
Two places. Real work at 4,000 companies: their documents, codebases, business records and workflows, with identities removed before delivery while amounts, dates, structure and outcomes stay intact. And Gerra's own builds: environments, tasks and evaluations written from scratch, such as synthetic spec-to-repo coding tasks graded by held-out assertion suites. Data collection runs across the United States, with supporting teams in India, the Philippines and Indonesia.
- Can Gerra build custom data or environments?
Yes. Gerra builds new collections, tasks, environments and evaluations around a lab's brief: the domain, tools, difficulty, reward shape and harness it needs. Existing environment families can also be extended with new variants.
- Who uses Gerra's data?
Frontier AI labs use Gerra's documents, codebases, RL environments and trajectories for pretraining, post-training, reinforcement learning and evaluation. Investment firms use its market data, such as retail investor sentiment, prediction-market probabilities and GPU rental prices. Gerra's case studies describe the work without naming the customer.
- Where can I buy RL environments for training AI agents?
Gerra builds runnable RL environments with machine graders. It has delivered 64,000 graded back-office tasks, 96,000 computer-use trajectories, 96,000 repository tasks, 9,600 spec-to-repo tasks, 16,000 exploit-and-patch security tasks, 12,800 browser-graded web tasks, 1,600 ML competition runs, 800 paper replications and 960 post-training runs. Each environment ships with its starting state, the task and its tools, a private grader and recorded agent runs.
- Where can I buy RL environments for training coding agents?
Gerra builds runnable coding environments graded by machine, not by a static answer key. Spec-to-repo tasks ask an agent to build a repository from a written spec and are graded by held-out assertion suites. Long-horizon repository tasks are mined from real repositories and checked as genuine fail-then-pass changes. Gerra has delivered 9,600 spec-to-repo tasks and 96,000 repository tasks, built against NL2Repo-Bench, SWE-bench Verified and SWE-bench Pro. Grading runs offline in a pinned container with the network off.
- What makes a task an RL environment rather than a benchmark?
Partial credit. A benchmark only says whether a finished attempt worked; an environment has to score partial progress, so the reward carries information. Gerra environments use three constructions: assertion fractions, where the score is the share of held-out assertions passed; subtask ladders, where a chained task reports the rung reached; and substep trees, where dependent steps are skipped rather than failed. Each environment also ships a reference solution that must score full marks and a stub that must score partial, so the grader itself is tested.
Partial Credit Is What Turns a Benchmark Into an Environment
- How do you prove a coding task can't be passed from memory?
Run the real library through your own grader and require it to fail. Gerra calls this the plagiarist gate: for every task with a public analogue, the genuine upstream package is installed into the task environment, scored by that task's own grader, and must fail before the task ships. It held on 8 of 8 gated tasks in the reference tranche. Crawl dates and obscurity prove nothing, because crawl date is not training cutoff and obscure code still gets crawled.
- What computer-use training data does Gerra have?
Gerra has delivered 96,000 browser and desktop trajectories. Each step records the screenshot and accessibility tree the agent saw, the grounded mouse or keyboard action it took, and the change that followed, and every trajectory ends with a final-state check against the application. Labs use them for supervised post-training and distillation, and the matching environments for reinforcement learning.
- What are Gerra's back-office RL environments?
Real, de-identified business operations turned into RL environments: accounting, reconciliation, billing, insurance and people workflows, each with a known correct answer graded to the cent. Gerra has delivered 64,000 graded back-office tasks. Each environment includes the connected company records, the tools the agent needs, a private grader and measured baseline runs.
- Does Gerra have environments for ML research agents?
Yes. Gerra's GPU-backed research environments cover ML competitions, rubric-graded paper replications and post-training runs that must beat the base model, each with recorded and graded agent runs. Gerra has delivered 1,600 competition runs, 800 paper replications and 960 post-training runs.
- Who sells complete codebases with full Git history for LLM training?
Gerra delivers full-history codebases as cloneable Git histories with source, tests, documentation, branches, tags, merges and reverts intact, one inspectable unit per repository. Across its sources Gerra holds over 6,000 codebases with about 1.4 million commits. Every bundle passes an integrity check before delivery.
- Does Gerra build security RL environments?
Yes. Gerra has delivered 16,000 exploit-and-patch tasks: containerized, multi-stage targets where an agent has to find a vulnerability, exploit it and patch it, with a machine grader for each stage, plus detect-and-patch pairs derived from real CVEs.
- Does Gerra have tasks for building whole applications?
Yes. Gerra has delivered 12,800 browser-graded web tasks from 9,600 app specifications. Each task describes a whole application in plain prose, a browser grades the running app, and substep trees give partial credit along the way.
- What professional documents does Gerra have?
Real work documents from engineering, finance, sales and consulting, delivered as the original files with their rendered pages, text, structure and a SHA-256 hash for each file. Gerra has processed 800,000 documents and processes hundreds of thousands more every week. Documents can come with the communication, project and operating records around them, and with human reference outputs for evaluation.
- Can company data be de-identified without destroying its training value?
Yes, if the removal is precise. Gerra's rule is to remove who it is and never what happened: identities become stable tokens while amounts, dates, sequence, structure and outcomes stay, because they carry the signal. The check has to read the shipped files rather than the build's own log, because automated checkers share assumptions with the scrubber and miss link targets, images, page-layer branding and per-recipient stamps.
Why Automated De-Identification Passes Corpora That Humans Reject
- Where can I get robot teleoperation data?
Gerra operates its own robot fleet and records demonstrations on it: 400K+ success-labeled episodes of people operating robots, with 35K+ new episodes a month. Each episode syncs RGB video, depth, joint states, IMU and audio to under a millisecond, across Booster T1, Unitree H1/G1 and LimX Tron robots. Data ships in the Open Robot Training Format and converts to LeRobot and RLDS.
- What data does Gerra build for investment firms?
Market and compute data built for point-in-time research: retail investor sentiment (426M+ messages since 2009), prediction-market probabilities mapped to 500+ securities, GPU rental prices (262,146 observations across 40 providers and 33 GPU SKUs), hosted-model inference prices, US federal software contracts mapped to 120+ tickers, and technographics on 10M+ companies. Each product page lists its coverage, formats and delivery.
- What should I check before buying alternative data?
Four things decide whether a backtest means anything. Whether the history is point-in-time or restated, since a rebuilt series shows a past no trader could have seen. Whether deleted records survive, since deletion is not random and survivorship makes a crowd look smarter than it was. Whether labels are human-applied or model-inferred, since an inferred label adds a second model error. And whether the vendor ran the backtest, since the vendor chose the universe, construction and period. Gerra publishes this checklist as a buyer guide that applies to any vendor, Gerra included.
How to Evaluate Retail Sentiment Data Before You Backtest It
- Does Gerra publish results when its own data doesn't work?
Yes. Gerra tested a four-year GPU rental panel walk-forward against the forward equity returns of exposed companies, found no standalone directional alpha at that panel size, period and feature set, and published the null. The GPU Rental Prices product is a separate, much denser record, 262,146 observations across 40 providers and 33 GPU SKUs, sold as compute-market intelligence for procurement, provider benchmarking and capacity planning, not as a trading signal.
- Can I see a sample first?
Yes. Every Gerra product page shows a real sample or takes a sample request, and public samples such as robot episodes and recorded agent runs are on the site. Samples are small; full deliveries are scoped with each customer. Requests go to team@gerra.com or through the sample form.
- How does Gerra deliver data, and in what formats?
It depends on the product. Market data ships over S3, REST and WebSocket in JSON, CSV and Parquet. Code and environments ship as Git bundles, Docker images and JSONL task indexes. Documents ship as the original DOCX and PDF files with JSON or Parquet structure. Robot data ships in the Open Robot Training Format with LeRobot and RLDS conversion. Every product page lists its delivery methods, formats and refresh cadence.
Anything else: team@gerra.com, or request a sample.