endless-learning

An open-source reproduction of knowledge-graph-driven RL task synthesis, as described in the Kimi K3 technical report (arXiv 2607.24653): a self-evolving concept hierarchy that agents expand through web exploration, from which verifiable RL training tasks are mined, grounded in pinned real-world materials, and gated through an explicit verification contract.

v0 scope: the data engine — graph construction, task mining, verification, and versioned datasets that ship with their reward functions and environments. No RL training loop; trainers own the rollout loop (docs/006).

Two task domains are live:

• Deep-search QA — questions grounded in a frozen, content-addressed corpus, gated V1–V5 (grounded / unique / non-trivial / solvable / stable), every pass rate stamped with the versioned environment it was measured in.
• Coding — SWE-bench-shaped instances mined from real repositories (PR mirroring: the commit's own test changes are the oracle), validated by execution against dual states (the oracle must fail on the broken image and pass on the fixed one), flake-screened per state, and offline replayable from pinned container images.

What is measured, and what is not

The design docs record the negative results where they happened; this is the short version, because it is the honest way to read everything above.

Difficulty measurement is expensive and noisy: per-task pass rate is ±0.24 at 5 rollouts (measured, and exactly binomial). Aggregate over ≥30 tasks; never read a single task's label.

This repo is the engine, not the dataset

Emitted datasets are outputs of running the pipeline, and they are not distributed here. A search-QA record cites content-addressed document hashes; a coding instance names a pinned container image. Without the corpus and the images, neither re-verifies — so shipping the JSONL alone would be shipping a claim with its evidence removed. endless dataset verify re-verifies against your corpus, built by your runs.

Making those artifacts portable — a corpus bundle carrying only the documents a dataset cites, and published images — is the prerequisite for releasing datasets at all.

Datasets compile into three consumer-validated interchange formats (docs/007): a verifiers taskset package, chorus TaskRecord registries, and Harbor task directories — each verified against the consumer's own code, not its documentation.

Quickstart

Build your own graph. Nothing here downloads ours — the package is the engine, and every artifact below is produced by your run, in your workspace.

pip install 'endless-learning[all]'   # or pick extras: [anthropic] [openai] [litellm] [ddg]

mkdir my-graph && cd my-graph
endless init                          # writes endless.toml + .endless/ here
endless kg expand --seeds knowledge --budget-nodes 0   # plant a seed pack, no keys needed
endless kg stats

Three seed packs ship with the package — knowledge, quantitative, earth-systems — and --seeds also takes a path, so your own pack is a TOML file with a seeds list. Seeding is local; growing the graph calls a model and a search backend, so set the credentials your endless.toml roles name and raise the budget:

export ANTHROPIC_API_KEY=...          # or the provider your endless.toml selects
endless kg expand --budget-nodes 50   # agents explore outward from the seeds

From there endless synth run mines search-QA tasks from the graph, and the endless coding commands onboard repositories into pinned images and mine executable instances — see The pipelines. Docker is required for the coding domain only.

Install (development)

pip install -e '.[dev]'      # extras: [anthropic] [openai] [litellm] [mcp] [all]
endless --help

The pipelines

# knowledge graph
endless init                                       # create a workspace here
endless kg expand --seeds knowledge                # grow the concept DAG
endless kg stats

# search-QA synthesis
endless synth run --n 20 --workers 6               # sample → write → gate → tasks.jsonl
endless dataset verify                              # offline re-verification (no network, no models)
endless dataset export-verifiers                    # verifiers-v1 taskset package

# coding domain
endless coding select --n 10                        # KG concepts → candidate repos (star-filtered)
endless coding onboard --repo <url>                 # pinned image + flake-screened test ledger
endless coding mine --since 2024-01-01              # PR-mirror candidates → dual-state validation
endless coding verify-instances                     # replay every instance from its pinned images
endless coding export-chorus                        # chorus TaskRecord registry
endless coding export-harbor                        # Harbor task directories
endless coding publish-images --registry <prefix>   # retag (push only with --push)

Design docs

The design lives in docs/ as a numbered series — start at docs/000-overview.md. Decisions and measured results (including the negative ones) are recorded where they happened.