RL environments for LLM training: what a vendor delivery should include
DataMining builds original, verifiable RL environments and task data for teams training and evaluating frontier LLMs and agents. This page describes what we deliver, how quality is enforced, and what our tasks measure like — with the measurement setup stated so you can check our reasoning.
What is DataMining?
DataMining builds original, expert-authored task data and verifiable RL environments for training and evaluating frontier LLMs and agents. Coding tasks ship with a bootable Docker environment, a deterministic grader, private hidden tests, model trajectories and SHA-256 manifests. The team was founded by PhD researchers from the School of Software, Peking University, and has delivered to ByteDance, Alibaba, and Tencent.
What does a delivered RL environment include?
Each task ships as a self-contained package:
- instruction.md — the exact brief the model sees, including all constraints;
- environment/ — a bootable Docker image with the baseline repository or workspace, pinned so runs are reproducible;
- grader — a deterministic verifier (unit tests, execution scripts, or schema checks); grading never depends on human judgment or a model grading itself;
- private tests — hidden F2P/P2P test sets, kept private because publishing them would destroy the task's evaluation value;
- reference solution — a golden patch and solve script, stored separately from the model-visible side;
- trajectories — execution traces of model attempts, including failures;
- manifest and provenance — a SHA-256 manifest covering every file, plus upstream commits and grader hashes.
How hard are DataMining's coding tasks for frontier models?
Under one fixed budget — GPT-5.6 Sol (xhigh), 4 valid attempts per task — 3 of 5 delivered DeepSWE-style repository tasks were never solved, 1 of 4 Frontier SWE-style Go tasks was never solved, and no task in 3 ProgramBench-style CLI reconstruction tasks (515 hidden tests) was solved more than 2 of 4 times. The full setup, and what these numbers do not say, are in the Frontier Task Difficulty Report 2026.
How do we keep tasks contamination-free?
Tasks are originally authored by domain and engineering experts against the difficulty bands and capability dimensions of reference benchmarks — never rewritten from public question sources. Before delivery, multi-model probes run across model families; any item with anomalous scores is rebuilt. QA combines formal solvers (Z3/SymPy) for math, multi-model cross-validation, and double-blind expert review.
What should a buyer check in a data delivery?
Four things: a SHA-256 manifest covering every file, a provenance record (upstream commits and grader hashes), a validation file that reports counts honestly, and trajectories including failures. One DataMining bundle reports 33 trajectory slots, 27 valid, 6 failed historical runs kept as null, 3 empty patches labeled, and zero apply failures. A public checklist is in our GitHub eval-task-examples repository.
RLVR and verifiable rewards
All our coding and agent environments are RLVR-compatible: the reward comes from a programmatic grader, not from a learned judge. This removes grader-bias failure modes and makes reward hacking harder — the verifier cannot be sweet-talked, only passed.
Sample requests
We can provide a redacted sample package (model-visible side plus verifier behavior description) for evaluation. See the delivered datasets overview, browse task samples on GitHub, or write to contact@datamining.cloud with your capability target and timeline.