skip to content
Reliquary
DECENTRALISED POST-TRAININGBITTENSOR SUBNET 81
RELIQUARY
Open app Request early access
OPEN SOURCE · BITTENSOR SUBNET 81

Decentralised
post-training on demand.

Turn your tasks into training data for your own model. Built on independent GPUs on Bittensor, with reinforcement learning and evaluation coming next.

Request early access See the workflow
Batch generation · early access. RL & evaluation · coming soon.
curl -fsS https://www.reliqua.ai/api/sft
Read-only task registry · no API key.Read the network guide
INFRASTRUCTURE VISUALIZATION
01 / PLATFORMExplore the workspace

Start with the task.

Build a dataset for your next model. An API and workspace for generation, with RL and evaluation coming next.

01 / TRAINING DATA APIEarly access

Batch generation

Turn your tasks into a private dataset. Inspect the examples and train with your own tools.

See the workflow
02 / REINFORCEMENT LEARNINGComing soon

RL on demand

Bring your model and a verifiable reward. Build training signal for the tasks that matter.

See the workflow
03 / MODEL EVALUATIONComing soon

Evaluation

Find what changed between checkpoints. Keep the answers and evidence behind each result.

See the workflow
Environment HubFind the tasks, output formats and graders for your workload.Open catalog
BATCH GENERATION / EARLY ACCESSTRAINING DATA API / TASK → DATASET

From a task to a dataset.

Generate examples for the tasks your model needs to learn. Keep inspectable JSONL data for your own post-training pipeline. The preview below does not submit a job.

  1. 01
    Define your task.

    Choose a supported model, environment and output format.

  2. 02
    Review your batch.

    Set the volume and token limits. Review the quote before confirming.

  3. 03
    Collect your data.

    Inspect the delivery and take compatible JSONL into your own pipeline.

WORKSPACE / EXAMPLE WORKFLOW
01 / Choose a task

Give the model a task.

Choose a model and a task environment. Define the examples you need.

Get early accessAvailable configurations depend on qualification.
Reliquarybatch / task
Task configurationRead-only example
01 / Model
Instruction model Illustrative selection
02 / Environment
Math reasoning Single-turn task
Output format
JSONL · prompt / response
EXAMPLE PROMPT

Explain why 17 is prime.

Math reasoning / single turn
MODEL / ENVIRONMENT / FORMATConfiguration preview
Example workflow · read-only illustration · no task is submitted
Explore batch generation Read the generation guide
Coming soonRL ON DEMAND

Train on the tasks that matter to you.

Bring your model and define what a good answer looks like. The planned workflow connects generated attempts, verifiable rewards and training checkpoints in one workspace, so you can inspect the signal behind each update.

Explore RL on demand Customer RL is coming soon. The SFT pilot is a separate workflow.
Find the learning signalIllustrative
Same task · different responses16 attempts / group

All fail

No reward contrast

Mixed

Useful reward contrast

All pass

No reward contrast
group_relativeContrast between outcomes
How to read the signal

Each dot is an example response: an outlined dot fails (reward 0), a filled dot passes (reward 1). Mixed rewards distinguish responses within a group. Identical rewards provide no group-relative contrast in this binary-reward example.

Illustrative groups, not live telemetry or the network group size. Reward contrast does not guarantee model improvement.
Coming soonEVALUATION

Know what changed. Keep the evidence.

Compare models on the tasks you care about. The planned reports bring scores, answers and grader details together, so you can trace a result back to the work behind it.

Explore evaluation Workflow illustration. No customer evaluation was executed.
From checkpoint to reportWorkflow concept
  1. 01
    Select a checkpointyour-model / checkpoint
  2. 02
    Choose a task suiteCode · Math · Logic
  3. 03
    Inspect the reportTask results, grader details and examples
report / exampleNot measured
Task
Deduplicate a list
Grader
Python task tests
Evidence
Model response + test output
Inspect example task
input:  [3, 1, 3]
expect: [3, 1]
Illustrative workflow · custom evaluation is coming soon. No evaluation was executed.
02 / LABEXPLORE THE LAB

We’re the lab scaling decentralised RL.

Independent operators contribute compute. Verified work builds the training signal. We publish the weights, methods and evidence — so you can build on the results.

RELEASED RESEARCH CHECKPOINTReliquary-4BQwen3-4B-Base · reinforcement learning

A 4B model trained with verifiable reward on math and code tasks. Open weights, published training methods and an inspectable evaluation setup. These are historical research results, not customer-service benchmarks.

Weights & methodology Read the mechanism
Published checkpoint comparisonSampled pass@1
Qwen3-4B-BaseReliquary-4B
Published sampled pass@1 percentages, Qwen3-4B-Base compared with Reliquary-4B.
BenchmarkPass@1 chartQwen3-4B-BaseReliquary-4B
MATH-50037.2%72.6%
HumanEval+52.1%72.9%
MBPP+43.7%67.6%
T=1.0 · 8 samples/problem · same evaluation setup
Pinned model card · single training run, not replicated.MATH-500 is near-domain. Published results, not a service guarantee.
BUILT IN THE OPENBittensorSubnet 81 · public networkHugging FaceOpen weights & model cardsGitHubOpen-source code & operator tools
RESEARCH NOTESThe reasoning behind the system.Mechanisms, experiments and release reports RELIQUARY-4B / TRAINING PAPERFrom training signal to open weights.Read the training paper

Public perspectives.

COMMENTARY & OPEN-SOURCE RESEARCH
ON THE POST-TRAINING MECHANISM
“one of the most compelling structures i have seen for a scalable post training mechanism on Bittensor”
const@const_reborn · September 30, 2026
Read the commentary
OPEN-SOURCE REFERENCE

Affine

Affine’s inference-verification rewrite cites Reliquary as inspiration.

“inspired by Reliquary (SN81)”
Inspect the sourceResearch direction
03 / NETWORKBITTENSOR / SUBNET 81

From independent compute
to published evidence.

Follow the work, not just the promise: task activity, verified data and published checkpoints. Open the source evidence behind each stage of Subnet 81’s learning loop. Public network activity is separate from customer orders.

Explore network activity
PUBLIC NETWORK SNAPSHOTExplore monitoring
Active tasks
—

Open entries in the public task registry.

Snapshot not loaded
Verified tokens
—

Observed SFT jobs, not customer dataset deliveries.

Snapshot not loaded
Published checkpoints
—

Current release per observed training run.

Snapshot not loaded
GENERATION RECORDS

Training signal, with evidence.

  1. Accepted—

    Admission checks recorded.

  2. Audited—

    An audit result was recorded.

  3. Proof passed—

    The declared proof passed.

Observed generation records appear when the public source is available.

Understand the evidence
TRAINING RECORDS

Models you can inspect.

Latest published checkpoint per observed run.

Open monitoring to inspect the available training records.

Explore training records

Shared public observations, not customer deliveries. Evidence stages are distinct; a proof does not establish trainer consumption. Source times and availability can differ. Open monitoring for current source detail.

How the subnet worksGENERATE / VERIFY / TRAIN / EVALUATE

Public network mechanism, separate from customer batches.

Prompt → rollout group

One task. Multiple attempts.

Independent operators generate responses to a task using the published checkpoint. Each group carries the tokens, rewards and proof needed to inspect the work.

Inspect network submissions
reliquary / subnet 81
01 / rollout groupREFERENCE
model
Qwen3.5-2B
tasks
math / code / logic
group size
8 rollouts
evidence
tokens · rewards · GRAIL proof

Protocol reference · not live telemetry

AROUND THE NETWORK

Built in a wider ecosystem.

Open research and compute providers in the Bittensor ecosystem. Each relationship is different.

AffineOpen-source research reference TargonGPU offers in our comparator LIUMGPU offers in our comparator ChutesBittensor inference ecosystem
04 / EARLY ACCESS Approved pilots

Need training data for your model or subnet?

Bring a representative task and your expected volume. We’re working with a small group of teams to shape the first workloads.

Request early access
Explore the workspace Contribute compute
your first workloadSCOPE / QUOTE / CONFIRM
01 / Model & task
The model, revision and environment you need
02 / Batch size
Prompt count, samples and token limits
03 / Acceptance criteria
Output format and how you’ll validate the data

No fixed public rate yet. We review compatibility and quote eligible workloads before confirmation.

BEFORE YOU START

A few practical questions.

What can I use today?

Environments, GPU comparison and network monitoring are public. Batch generation requires approved early access and a qualified configuration. RL on demand and custom evaluation are coming soon. The SFT pilot is separate; managed GPU rental is not publicly available.

What do I submit and receive?

Bring a task, a supported model and environment, and your expected volume. Approved orders provide delivery files and generation details. Compatible, validated JSONL can be saved to Datasets or used in your own pipeline.

How is a workload priced?

Model, task, volume and output-token limits define the scope. There is no fixed public rate yet. Eligible configurations receive a quote before confirmation; an estimate is not a delivery guarantee.

How do I assess the data quality?

Set validation requirements before generation. Review a representative sample against your acceptance criteria. Network verification checks submitted work; it does not guarantee suitability for every training task.

Is this the same as the live subnet training?

No. Public Subnet 81 activity is the network’s own generation and reinforcement-learning loop. Customer batches, private datasets and training runs have separate access and execution requirements.