Inserloft / Official documentation

Everything Inserloft publishes.

The complete developer reference for Inserloft models, datasets, benchmarks, research artifacts, source code, APIs and public interfaces. Every resource, every endpoint, every integration path — documented in one place.

Quickstart Ecosystem map API status

ModelsParrot Coder · 200M params · Apache 2.0

DatasetsPyMini · PyMAX · ChatESP

BenchmarksRockman v0.2 · 220 tasks

ResearchProteus Structures · PDB artifacts

Source code3 public repositories

Hosted APINot yet public

Local interfaceOllama HTTP · /api/generate · /api/chat

Ecosystem map

Every public Inserloft surface, where it lives, and what it exposes. inserloft.com ↗

Surface	Type	Location	Status
Parrot Coder	Model	Hugging Face / Ollama	Available
PyMini	Dataset	Hugging Face	Available
PyMAX	Dataset (gated)	Hugging Face	Available
ChatESP	Dataset	Hugging Face	Available
Rockman	Benchmark	GitHub / Hugging Face	Available
Proteus Structures	Research artifacts	Hugging Face	Available
Rockman	Source code	GitHub	Public · MIT
docs	Source code	GitHub	Public · TypeScript
.github	Metadata	GitHub	Public
Ollama HTTP	Local API	Local machine	Available
Hugging Face API	Hosted API	huggingface.co	Available (HF)
GitHub API	Hosted API	api.github.com	Available (GitHub)
Hosted Inserloft API	Hosted API	—	Not public

Quickstart

Run Parrot Coder locally through Transformers, Ollama, or directly from Hugging Face. No hosted Inserloft API is required. Open on Ollama ↗

Parrot Coder Transformers

from transformers import AutoModelForCausalLM, AutoTokenizer

model_id = "InserloftResearch/Parrot-Coder"
tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(model_id)
prompt = "def fibonacci(n):"
inputs = tokenizer(prompt, return_tensors="pt")
outputs = model.generate(**inputs, max_new_tokens=50)
print(tokenizer.decode(outputs[0], skip_special_tokens=True))

Parrot Coder Ollama

ollama run Inserloft/ParrotCode

# Or via HTTP
curl http://localhost:11434/api/generate -d '{"model":"Inserloft/ParrotCode","prompt":"Write a Python function","stream":false}'

PyMini Hugging Face Datasets

from datasets import load_dataset

ds = load_dataset("Inserloft/PyMini")
print(ds["train"][0])

Interface scope: these examples use published model distributions and Ollama's local runtime. There is no public hosted Inserloft API represented here.

Parrot Coder

Inserloft's first open-source model. A highly efficient, 200M parameter language model optimized for rapid code autocompletion and code generation tasks. Hugging Face ↗

~200M params 2K context FP16 GGUF 517 MB Ollama Apache 2.0 Safetensors GPT-2 base + LoRA 428 downloads/month

Direct inference Python

from transformers import AutoModelForCausalLM, AutoTokenizer

model_id = "InserloftResearch/Parrot-Coder"
tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(model_id)
prompt = "def fibonacci(n):"
inputs = tokenizer(prompt, return_tensors="pt")
outputs = model.generate(**inputs, max_new_tokens=50)
print(tokenizer.decode(outputs[0], skip_special_tokens=True))

Model details: Parrot Coder is built with GPT-2 as the base architecture and merged LoRA adapters. It ships as both a merged Safetensors checkpoint (0.3B params) and an FP16 GGUF model for local deployment. For GGUF versions, visit the Ollama Registry.

Model registry

All public Inserloft models across distribution platforms.

Model	Params	Base	Format	License	Platform
Parrot-Coder	~200M	GPT-2 + LoRA	Safetensors / GGUF	Apache 2.0	HF · Ollama

Roadmap: Inserloft has announced additional model families for reasoning, everyday intelligence, analysis, coding and visual perception. No additional public releases are available at this time.

PyMini

Compact synthetic Python instruction dataset. 49,202 rows with examples in English and Spanish. Hugging Face ↗

49.2k rows 6 instruction classes EN / ES MIT Parquet Updated Jul 2026

Load dataset Python

from datasets import load_dataset

ds = load_dataset("Inserloft/PyMini")
print(ds)

# DatasetDict({ train: Dataset({ features: [...], num_rows: 49202 }) })

# Access a sample
print(ds["train"][0])

Sample rows JSON

{
  "instruction": "The following code contains an error. Fix it.",
  "input": "if x > 0\n    print('ok')",
  "output": "if x > 0:\n    print('ok')"
}
{
  "instruction": "What is the output of the following code?",
  "input": "print([x*2 for x in range(3)])",
  "output": "[0, 2, 4]"
}

Filter short outputs Python

from datasets import load_dataset

ds = load_dataset("Inserloft/PyMini", split="train")
short = ds.filter(lambda r: len(r["output"]) < 200)
print(short.num_rows)

PyMAX

Large-scale gated Python dataset. 500M examples across 100 Parquet shards with code-quality and complexity metadata. Hugging Face ↗

500M examples 100 shards Access controlled Parquet Updated Aug 2026

Access controlled: PyMAX requires an accepted access request on Hugging Face. Once granted, you can load it with the standard datasets library.

Load gated dataset Python

from datasets import load_dataset

# Requires prior access approval on Hugging Face
ds = load_dataset("InserloftResearch/PyMAX", streaming=True)
for row in ds["train"]:
    print(row)
    break

ChatESP

Spanish-language conversational dataset published through Inserloft Research. Hugging Face ↗

Hugging Face Spanish Conversational Research

Load dataset Python

from datasets import load_dataset

ds = load_dataset("InserloftResearch/ChatESP")
print(ds["train"][0])

Rockman v0.2

Contamination-resistant benchmark for evaluating LLMs on executable software-engineering tasks. GitHub ↗ · Hugging Face ↗

220 tasks Pass@1 2 s timeout 256 MB memory Temperature 0 MIT license Python

Rockman evaluates executable software-engineering tasks instead of textual similarity. Its documented task families include generation, debugging, optimization, stateful systems and multi-step workflows. The benchmark is designed to resist training-data contamination through six integrity layers.

Installation Shell

git clone https://github.com/InserLoft/Rockman
cd Rockman
pip install -e .

Load from Hugging Face Python

from datasets import load_dataset

ds = load_dataset("Inserloft/Rockman")
print(ds)

Evaluation splits

Public, private and hidden task splits with their purposes.

Split	Tasks	Tests	Purpose
Public	154	Public	Development and local experimentation
Private	44	Hidden	Primary official evaluation split
Hidden	22	Hidden	Additional leaderboard integrity

Implementation stack

Public implementation details of Rockman v0.2.

Area	Public implementation
Evaluator	Python package under rockman/
Runners	Python, C++, Java, JavaScript, TypeScript, Rust, Go
Sandbox	Docker + local fallback
Security	AES-256-GCM hidden-test encryption
Metrics	Pass@1, difficulty/category/task-type breakdowns, Wilson intervals
Integrity	Canary strings, SHA-256 hashing, split isolation, source-attribution metadata, temporal controls

Project Proteus

AI-assisted biological discovery, protein intelligence, generative biology and computational research. Proteus Structures ↗

Public artifacts

Structure files

Public structure files are distributed as PDB artifacts for computational inspection. Predicted structures are explicitly distinguished from experimentally validated results.

CNB-1First experimental model entry in the public Proteus collection, described as Caelis Neural Base-1.

Macromolecular_PET_Hydrolase_Alpha.pdb~350 aa computational candidate associated with a PET-hydrolysis hypothesis requiring experimental validation.

Alanine-Glycine-Aspartate.pdb377 aa predicted structure; the public card notes that experimental structure and biological function are not established.

Inspection tools

Work directly with PDB files

Public structures can be inspected with standard molecular-visualization tools and downloaded from the Hugging Face repository.

PDBPyMOLChimeraXVMDMol*3Dmol.js

Open structures ↗ See artifacts ↗

Research caution: these artifacts are computational objects. Do not read a predicted structure as proof of biological activity or an experimentally confirmed mechanism.

Proteus file reference

Download and inspection details for each public structure artifact.

File	Type	Residues	Status	Download
CNB-1	PDB	—	Experimental model entry	HF
Macromolecular_PET_Hydrolase_Alpha.pdb	PDB	~350 aa	Computational candidate	HF
Alanine-Glycine-Aspartate.pdb	PDB	377 aa	Predicted structure	HF

Ollama HTTP API

ParrotCode is distributed through the Inserloft Ollama namespace. Ollama exposes its standard local HTTP server on your machine. Ollama Registry ↗

Available endpoints

Local HTTP server

Standard Ollama API on port 11434. No authentication required for local use.

POST/api/generateText generation

POST/api/chatChat completion

GET/api/tagsList models

POST/api/pullPull model

DELETE/api/deleteDelete model

GET/api/versionVersion

Open Ollama ↗

Base URL

http://localhost:11434

Default Ollama port. All endpoints accept JSON payloads and return JSON responses.

Available locally

Model name

Inserloft/ParrotCode

FP16 GGUF · 517 MB · 2K context

Generate via cURL Shell

curl http://localhost:11434/api/generate \
  -d '{
    "model": "Inserloft/ParrotCode",
    "prompt": "Write a Python function to reverse a string",
    "stream": false
  }'

Chat via cURL Shell

curl http://localhost:11434/api/chat \
  -d '{
    "model": "Inserloft/ParrotCode",
    "messages": [
      {"role": "user", "content": "Explain decorators in Python"}
    ],
    "stream": false
  }'

Hugging Face API

All Inserloft models and datasets are hosted on Hugging Face. Use the standard HF APIs to list, download, and inspect them. Inserloft Research ↗

Available endpoints

Hugging Face Hub API

REST API for browsing Inserloft's public models and datasets.

GET/api/models?author=InserloftResearch&limit=100List models

GET/api/models/InserloftResearch/Parrot-CoderModel info

GET/api/datasets?author=InserloftResearch&limit=100List datasets

GET/api/datasets/Inserloft/PyMiniDataset info

GET/api/datasets/InserloftResearch/PyMAXGated dataset

GET/api/datasets/InserloftResearch/ChatESPDataset info

GET/api/datasets/Inserloft/RockmanBenchmark data

GET/api/models/InserloftResearch/Proteus-StructuresArtifacts

Open Hugging Face ↗

Base URL

https://huggingface.co

Standard HF Hub API. No authentication required for public resources. Gated datasets require a token.

Available via HF

List models & datasets cURL

curl "https://huggingface.co/api/models?author=InserloftResearch&limit=100" | jq '.[].modelId'
curl "https://huggingface.co/api/datasets?author=InserloftResearch&limit=100" | jq '.[].id'

Query model / dataset info cURL

curl https://huggingface.co/api/models/InserloftResearch/Parrot-Coder
curl https://huggingface.co/api/datasets/Inserloft/PyMini

PyMini — Datasets Server cURL

curl -X GET "https://datasets-server.huggingface.co/rows?dataset=Inserloft%2FPyMini&config=default&split=train&offset=0&length=100"
curl -X GET "https://datasets-server.huggingface.co/splits?dataset=Inserloft%2FPyMini"
curl -X GET "https://huggingface.co/api/datasets/Inserloft/PyMini/parquet/default/train"

ChatESP — Datasets Server cURL

curl -X GET "https://datasets-server.huggingface.co/rows?dataset=InserloftResearch%2FChatESP&config=default&split=train&offset=0&length=100"
curl -X GET "https://datasets-server.huggingface.co/splits?dataset=InserloftResearch%2FChatESP"
curl -X GET "https://huggingface.co/api/datasets/InserloftResearch/ChatESP/parquet/default/train"

PyMAX — Datasets Server (gated) cURL

curl -X GET -H "Authorization: Bearer $HF_TOKEN" \
  "https://datasets-server.huggingface.co/rows?dataset=InserloftResearch%2FPyMAX&config=default&split=train&offset=0&length=100"
curl -X GET -H "Authorization: Bearer $HF_TOKEN" \
  "https://datasets-server.huggingface.co/splits?dataset=InserloftResearch%2FPyMAX"
curl -X GET -H "Authorization: Bearer $HF_TOKEN" \
  "https://huggingface.co/api/datasets/InserloftResearch/PyMAX/parquet/default/train"

Datasets Server: The datasets-server.huggingface.co endpoints return paginated rows, split metadata and parquet file lists. Public datasets need no token; PyMAX requires HF_TOKEN.

GitHub API

Access Inserloft's public repositories, releases, issues and metadata through the GitHub REST API. GitHub organization ↗

Available endpoints

GitHub REST API v3

Standard GitHub API. No authentication required for public repositories.

GET/orgs/InserLoft/reposList repositories

GET/repos/InserLoft/RockmanRepo metadata

GET/repos/InserLoft/Rockman/releasesReleases

GET/repos/InserLoft/Rockman/contentsFile tree

GET/repos/InserLoft/docsDocs repo

GET/repos/InserLoft/.githubOrg metadata

Open GitHub ↗

Base URL

https://api.github.com

Standard GitHub REST API v3. Rate-limited to 60 requests/hour without authentication.

Available via GitHub

List repositories cURL

curl https://api.github.com/orgs/InserLoft/repos
curl https://api.github.com/repos/InserLoft/Rockman/releases/latest

Local inference

Run Parrot Coder entirely offline with Transformers or llama.cpp. No network required after download.

Transformers Python

from transformers import AutoModelForCausalLM, AutoTokenizer
import torch

model_id = "InserloftResearch/Parrot-Coder"
tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(
    model_id,
    torch_dtype=torch.float16,
    device_map="auto"
)
prompt = "def quicksort(arr):"
inputs = tokenizer(prompt, return_tensors="pt").to(model.device)
outputs = model.generate(**inputs, max_new_tokens=100)
print(tokenizer.decode(outputs[0], skip_special_tokens=True))

llama.cpp Shell

# Download GGUF from Ollama
ollama pull Inserloft/ParrotCode

# Run directly with llama.cpp
./main -m parrot-coder.gguf -p "def hello():" -n 64

API status matrix

Complete status of every Inserloft-related API and interface.

API / Interface	Status	Base URL	Auth
Ollama HTTP	Available	http://localhost:11434	None
Hugging Face Hub API	Available	https://huggingface.co	None / Token (gated)
GitHub REST API	Available	https://api.github.com	None / Token
Hugging Face Datasets	Available	datasets library	None / Token (gated)
Hosted Inserloft API	Not public	—	—
Inserloft SDK	Not public	—	—

Repositories

All publicly visible Inserloft repositories. GitHub organization ↗

Repository	Visibility	Language	License	What it exposes	Open
InserLoft/Rockman	Public	Python	MIT	Benchmark implementation, releases, task sets, CLI and evaluation infrastructure	GitHub ↗
InserLoft/docs	Public	TypeScript	—	Docs source: React 19, Vite 7, TypeScript 5.9, Tailwind 4, marked, highlight.js, lucide-react	GitHub ↗
InserLoft/.github	Public	—	—	Organization-level GitHub metadata	GitHub ↗

CLI reference

Command-line tools available for Inserloft resources.

Command	Tool	Purpose
ollama run Inserloft/ParrotCode	Ollama CLI	Run ParrotCode interactively
ollama pull Inserloft/ParrotCode	Ollama CLI	Download the GGUF model
git clone https://github.com/InserLoft/Rockman	Git	Clone Rockman benchmark
pip install -e .	pip	Install Rockman locally
huggingface-cli download InserloftResearch/Parrot-Coder	HF CLI	Download model files
huggingface-cli download Inserloft/PyMini --repo-type dataset	HF CLI	Download dataset

Status matrix

Current state of each public surface and its primary distribution location.

Surface	Status	Examples	Primary location
Models	Available	Parrot Coder (200M)	Hugging Face
Datasets	Available	PyMini (49.2k) · PyMAX (500M) · ChatESP	Hugging Face
Benchmark	Available	Rockman v0.2 (220 tasks)	GitHub
Research artifacts	Available	Proteus Structures (PDB)	Hugging Face
Source code	Available	Rockman · docs · .github	GitHub
Local HTTP	Available	Ollama /api/generate · /api/chat	Ollama
Hosted API / SDK	Not public	No verified public endpoint	—

Source discipline: this page only surfaces resources that are publicly verifiable. Internal or private projects are not represented as developer resources.

External surfaces

All Inserloft-related locations outside of inserloft.dev.

Platform	What lives there	Open
inserloft.com	Main product site, model registry, artifacts	inserloft.com ↗
Hugging Face (Research)	Parrot Coder, PyMAX, ChatESP, Proteus Structures	HF Research ↗
Hugging Face (Core)	PyMini, Rockman dataset	HF Inserloft ↗
GitHub	Rockman, docs, .github	GitHub ↗
Ollama	ParrotCode GGUF distribution	Ollama ↗
News	Inserloft news and updates	News ↗

Changelog

Recent additions and updates across the Inserloft public surface.

Sep 2026Rockman repository updated on GitHub. Parrot Coder updated on Hugging Face.

Aug 2026PyMAX dataset published on Hugging Face (500M examples, gated).

Jul 2026PyMini dataset updated (49.2k rows). Parrot Coder model released on Hugging Face. Ollama registry entry published.

Jun 2026.github repository updated. GitHub organization metadata refreshed.

EarlierProteus Structures, ChatESP, and initial Rockman releases.