Every arena on the network, as the files the definitions are built from. The pages here are the same ones the site shows at microtensor.cloud/network. Hidden items never live here; validators receive them only inside encrypted epoch bundles.
| Arena | Task | Envelope | Network |
|---|---|---|---|
| ctx_compress | Agent context compression | mt-5 | testnet |
| doc_parse | Documents: page parsing | mt-5 | testnet |
| sql_gen | Databases: SQL generation | mt-5 | testnet |
Each folder holds:
| File | What it is |
|---|---|
arena.yaml |
The arena: purpose, specialised SLM, evaluation, envelope group, scoring, harness, router, escalation, floors, item counts, and the evaluation block (allowed hardware profiles, images by digest, probe, spot check, ban and canary numbers) |
schema/input.json, schema/output.json |
The task's input and output schemas |
ARENA.md |
The arena page, rendered from arena.yaml |
data/README.md |
The attached dataset and how to build dev items from it |
CHANGELOG.md |
What changed and when |
| File | What it is |
|---|---|
shared/envelopes.yaml |
The six envelope classes (mt-1 Budget phone, mt-2 Phone, mt-3 CPU, mt-4 AI PC, mt-5 Small GPU, mt-6 Large), which are open on testnet and mainnet, their rules and how validators measure them |
shared/bases.yaml |
Every allowed base model with its pinned revision, licence, size, the groups it fits and any licence condition |
shared/profiles.yaml |
The hardware profiles miners and validators run on (P1 to P3, K, C): one GPU architecture, precision and batching mode each, with the calibration rule a profile must pass |
An arena names one group and may only tighten its limits. It names exactly one base model, the one miners fine tune from and are scored against; it comes from the shared list and must fit its group. An arena lists its allowed profiles from the shared list and pins every image by digest, never by tag alone. The same values are served at GET /v2/envelopes, GET /v2/bases and GET /v2/profiles.
pip install -e ".[compress]"
mta dev --arena ctx_compress --count 200 --out data/dev
mta check data/dev
mta dev builds items in the arena's exact format from its public dataset: the compaction instruction with its token budget, the session inside <session> tags, and the label the grader scores against (key facts with their aliases and probe questions with their answers). Twins are another cut of the same session. Hidden items are built the same way from fresh sessions that are never published, so train on the skill, not on these items. Pass --items data/dev/items.jsonl to mt envelope check to time your model on them.
mta render writes each ARENA.md and data note from its arena.yaml; mta render --check fails when one is out of date.
Source available under the Microtensor licence in LICENSE. Datasets keep their own licences.