Relearn

Public challenge repo for the Cortex Relearn subnet (challenge id relearn).

This repo is miner + eval only. It does not contain the Cortex control plane, gateway, or validator. Those live in CortexLM/cortex.

Pin	Value
Base model	Qwen/Qwen3.8-27B (Apache-2.0, native VLM)
Teacher / judge	HTTP API, wire id glm-5.3 (operator sets RELEARN_TEACHER_*)
Teacher weights	incoai/GLM-5.3-NVFP4, served from RELEARN_TEACHER_LOCAL_DIR
Eval image	ghcr.io/cortexlm/relearn-eval, pinned by digest
Score	Displacement vs the previous champion on a private holdout
Trust	No TDX / no Phala CVM. Miner pays Lium. The holdout reaches the pod only for the length of a run

The pin of record is config/relearn-pin.toml in the control plane: it carries this repo's git SHA and the eval image digest. Deploy = bump the pin after this repo's CI publishes a green image.

While eval_image_digest is empty, live hosts answer 503 on POST /v1/submissions rather than scoring — there is no simulated fallback.

Eval image

The image is the whole live scorer. The control plane boots it on a digest pin, stages the run request into /tmp/relearn_eval, and accepts a score only when the pod prints RELEARN_METRICS=<document> and RELEARN_EVAL_OK.

docker build -f eval/Dockerfile.scoring -t relearn-eval:dev .
relearn-eval score --request request.json --out metrics.json

Full contract, environment, and operator notes: docs/EVAL-IMAGE.md. Normative source: docs/RELEARN.md § Eval image contract in the control plane.

Miner path

See docs/ and the control-plane mirror docs/external-miner/relearn.md.

curl -sS -X POST https://<gateway>/challenge/relearn/v1/submissions \
  -H 'content-type: application/json' \
  -H "X-Lium-Api-Key: $LIUM_API_KEY" \
  -d '{
    "miner_hotkey": "<64-hex>",
    "artifact_digest": "<sha256>",
    "manifest": {"train_item_ids": [1], "train_dataset_ids": ["my-sft-mix-v3"]}
  }'

manifest is required evidence: an undeclared one fails the contamination gate rather than skipping it. Never commit LIUM_API_KEY or any secret.

Layout

Path	Role
eval/Dockerfile.scoring, eval/bin/relearn-eval	The digest-pinned CUDA scoring image
eval/src/relearn_eval/contract.py	Markers, schema, and the metrics document
eval/src/relearn_eval/request.py	The harvest request, and what it refuses
eval/src/relearn_eval/scoring.py	Every series in the document, measured
eval/src/relearn_eval/verify.py, harvest.py	The control plane's acceptance checks, mirrored
eval/src/relearn_eval/teacher.py	Frozen judge over HTTP; never serves miner weights
eval/src/relearn_eval/catalog/	Public, canary, general-canary, agent-trace slices
eval/src/relearn_eval/generators.py, decontam.py	Miner-side disjoint factory and bench blocklist
tests/	Image contract tests
docs/	Miner-facing and operator-facing docs

Teacher serving

The teacher is judge-only, and the scored artifact is always the miner weights loaded inside the eval image — never anything served through the teacher API. Download the NVFP4 weights, then point vLLM at RELEARN_TEACHER_LOCAL_DIR; never pass the Hugging Face repo id to vLLM. GPU shape is the operator's, described in the control-plane pin.