LoopNet

Ground truth for self-improving systems.

Structured loop designs, execution trajectories, outcomes, and failure modes — so you can train, evaluate, and debug loops with data, not anecdotes.

Load the dataset · End-to-end tutorial · Contribute records · Data card

---

Why this exists

Computer vision had ImageNet. RL had MuJoCo. Loop engineering had no shared corpus.

LoopNet fills that gap: every record is a complete loop story — spec, trajectory, outcome, LES breakdown, and when things break, a fail.* code from the shared taxonomy.

---

What you can do with it

Use case	How LoopNet helps
Failure prediction	42% labeled failures — models learn what breaking looks like
Benchmark generalization	Same schema as LoopBench holdout (v0.2)
Zero-cost replay	Feed LoopGym ReplayEnv — no API spend
Research & fine-tuning	JSONL + Parquet + Hugging Face — train on loop structure, not chat logs
Community contributions	ln/record-v1 + submission guide

---

Corpus at a glance (v0.2)

Records	545 (500 seed + 45 captured LoopGym runs)
Failure rate	40% — meets corpus policy
Schema	ln/record-v1 · pins lss@1.0.0 + les@1.0.0
Source	Synthetic seed + SimEnv captures (3 LoopBench envs)
License	Code MIT · Dataset CC BY 4.0

Seed-only v0.1 (500 records) remains at loopnet-seed-v0.1.

---

Load in one minute

Hugging Face (recommended — v0.2):

from datasets import load_dataset

ds = load_dataset("KanakMalpani/loopnet-v0.2", split="train")
print(ds[0]["outcome"], ds[0]["pattern_slug"])

Seed-only v0.1:

ds = load_dataset("KanakMalpani/loopnet-seed-v0.1", split="train")

Stream from GitHub (no clone):

ds = load_dataset(
    "json",
    data_files="https://raw.githubusercontent.com/KanakMalpani/loopnet/main/data/seed/records.jsonl",
    split="train",
)

Replay in LoopGym:

import loopgym as lg

env = lg.make("replay/loopnet-v1")
obs = env.reset(record_id="ln-00042")  # trajectory from corpus

---

Where it sits

flowchart LR
  CORE[Loop Core Engineering]
  NET["<b>LoopNet</b><br/>you are here"]
  GYM[LoopGym ReplayEnv]
  BENCH[LoopBench holdout]

  CORE --> NET
  NET --> GYM
  NET -.-> BENCH

Loading

Layer	Repo
Specs & failure codes	Loop Core Engineering
Dataset	LoopNet
Execution	LoopGym
Observability	loop-observability
Public scores	LoopBench

---

Repository map

Path	Purpose
schema/loopnet-record-v1.json	Canonical record schema
data/seed/records.jsonl	Seed corpus
scripts/validate_record.py	Schema + policy validation
scripts/generate_seed.py	Deterministic regeneration (--seed 42)
guides/COMMUNITY-SUBMISSION.md	Contribute records via PR
submissions/community/	Community JSONL inbox

---

Citation

@dataset{loopnet_v02,
  title={LoopNet v0.2},
  author={Malpani, Kanak},
  year={2026},
  url={https://huggingface.co/datasets/KanakMalpani/loopnet-v0.2}
}

_{Contributing · Security · Status}