Your traces and code, turned into benchmarks.
You already have the raw material: logs of what your models and agents did, and the codebases they work in. Nenyax shapes it into benchmarks, RL environments and training datasets, checked before they count.
Four kinds of output.
A benchmark measures a model. An RL environment trains one. A dataset feeds supervised or preference training. All of them are checked before they count.
GSM8K200 questions · seed 0
Qwen2.5-1.5B14.0%
SmolLM3-3B9.5% Qwen2.5-0.5B5.0%Real prompts a checker can score
Turn the requests your app actually saw into tasks with a verifier, so you measure models on your own work.
nenyax checkreasoning_gym:basic_arithmetic
A world a model can practise in
Tasks, a verifier and a reward, checked through four tiers before it counts as trainable.
nenyax traces import claude-code \ --since 7d --limit 500scrubbed secrets and personal data removed
SFT, DPO and distillation sets
Prompt and answer pairs from your best traces, chosen and rejected pairs from scores and feedback, or a strong model's answers for a smaller one to learn from.
Private by default
Everything you push starts private. Make it public and it joins the hub under your name.
Every call your app made is a lesson.
A trace is the record of one call to a model: the prompt, the answer, and any score or feedback it got. Send them from any OpenTelemetry SDK, import them from the tools you use, or import your coding-agent sessions.
Import from where they liveLangfuse, LangSmith, Phoenix, Braintrust, W&B Weave and Helicone, with their scores and feedback.
Or your coding agentArkenOS, Claude Code, Codex, Cursor, opencode and Pi sessions, straight from your machine.- Scrubbed firstThe CLI removes secrets and personal data on your machine before anything is uploaded.
Langfuse
LangSmith
Phoenix Claude Code sessionsBring traces from the tools you already use.
LangfuseTraces and scores from a projectLangSmithRuns and feedback from a projectPhoenixSpans from Arize Phoenix
BraintrustLogs and scores from a project
W&B WeaveCalls and feedback from a project
HeliconeRequests logged through Helicone
ArkenOSCoding-agent sessionsClaude CodeCoding-agent sessions
OpenTelemetryOpenLLMetry, OpenInference, LogfireYour coding agent builds the environment.
Install the build skill in your repository and describe the task. Your agent writes the tasks, the verifier and the reward; Nenyax audits them before anything is pushed.
- ArkenOS, Claude Code, Codex or CursorWhichever agent you already use, with the same skill.
- Audited before it countsLoads, completes an episode, reproduces from a seed, and the verifier tells a good answer from a bad one.
- Private until you publishNew listings start private. Make one public and it joins the hub under your name.
Build with your agentnenyax skill install
ArkenOS Claude Code
Codex
CursorWhat each output is for.
| Measures a model | Trains a model | Needs a verifier | |
|---|---|---|---|
| Benchmark tasksreal prompts a checker can score | |||
| RL environmenttasks, verifier and reward | |||
| SFT pairsprompt and answer from your best traces | |||
| DPO preferenceschosen and rejected answers | |||
| Distillation seta strong model's answers |
Four steps from raw material to a benchmark.
Bring it in
Import traces or install the skill in a codebase.
Pick what to make
Benchmark tasks, an RL environment, or a dataset.
Audit it
Every push runs the same checks; essentials can't be skipped.
Use or publish
Evaluate and train on it, or share it in the hub.
Three commands.
The CLI scrubs secrets and personal data before anything is uploaded.
Good to know.
nenyax traces import claude-code --since 7d nenyax push ./my-env --privateprivate until you publish