Protomire
A dedicated IDE for the complete model lifecycle, built on Code – OSS.
MODEL DEVELOPMENTExplore project ↗
HOME OF BACKWOODS MODELS & PROTOMIRE
Frontier? Nah.
We’re out in
the backwoods.
Welcome to the home of Backwoods Models and Protomire. Models, tools, and experiments from the other end of the frontier. Built by Greg Havens.
BUILT IN THE OPEN
The opposite of frontier. The same curiosity.
↓ KEEP EXPLORING01 / BACKWOODS MODELS
Small clearing.
Big experiments.
Welcome to the backwoods—the opposite of frontier. A home for hands-on model building, fine-tuning, distillation, and the datasets that make it possible.
Explore Fabletron and the published model builds, dig into coding traces, or meet Biff: the next model taking shape out here.
Explore the models ↗ON THE HORIZON THE NEXT PROJECT
Meet Biff.
A model that makes decisions.
Biff is an upcoming open-weight System One model. Give it context and typed questions. Get probabilities over the answers you define—ready for your code to act on.
Built for judgments, routing, and scoring, without generating an answer you have to parse.
In development · Weights not yet released
Planned release: Hugging Face · Apache 2.0
ILLUSTRATIVE OUTPUT · NOT LIVE INFERENCE
02 / THE HOME OF PROTOMIRE
One model.
Its whole lifecycle.
Protomire is a model-development IDE built on Code – OSS. Bring research, datasets, training recipes, evaluation, and exports together in one project.
A native MCP server gives agents access to the same operations as the human workbench. Build your recipe, track the runs, and keep the model’s history connected.
Research & dataTraining & evaluationRecipes & lineageNative MCP
Explore Protomire ↗In active development. See the repository for verified capabilities and platform delivery status.
THE OPEN WORKBENCH
Built to be used.
Shared to be improved.
Explore the code, the models, and the datasets.
Choose a project. Get the whole story.
CODE, WEIGHTS & DATA ↙
Triggerfish
A multi-channel agent platform with policy enforcement below the LLM layer.
AI AGENT PLATFORMExplore project ↗
Visit trigger.fish ↗Moonshiner
From seed tasks to judged model traces and training-ready datasets.
TRAINING DATAExplore project ↗
VCF Agent Skills
28 version-aware agent skills for VCF 9.0, 9.1, and 9.1.1.
VMWARE CLOUD FOUNDATIONExplore project ↗
Fabletron
A Nemotron fine-tune for agentic coding and tool use, with evaluation tooling.
MODEL & EVALUATIONExplore project ↗
Jev Graph Builder
Turn documents into a searchable knowledge graph with checked relationships.
KNOWLEDGE GRAPHSExplore project ↗
Jev No Bullshit
Check an assistant’s final claims against the work it actually performed.
AGENT RELIABILITYExplore project ↗
Jev Scope Control
Check whether a proposed tool call stays within the agreed scope.
AGENT GUARDRAILSExplore project ↗
The Reef
The public skill and plugin registry for Triggerfish.
SKILLS & PLUGINSExplore project ↗
Triggerfish CMS Auth
GitHub OAuth endpoints for Decap CMS on Cloudflare Pages.
WEB INFRASTRUCTUREExplore project ↗
ESXi Data Collection
Collect hardware and configuration details from ESXi hosts over SSH.
INFRASTRUCTUREExplore project ↗
Fabletron · LoRA
The fine-tuning adapter behind the Fabletron model family.
PEFT ADAPTERExplore project ↗
Fabletron · GGUF
A GGUF export for running Fabletron with compatible local runtimes.
LLAMA.CPP / LM STUDIOExplore project ↗
Fabletron · NVFP4
Fabletron in NVIDIA ModelOpt FP4 format for Blackwell serving.
NVIDIA BLACKWELLExplore project ↗
Fabletron · MLX
A four-bit MLX export of Fabletron for Apple Silicon workflows.
APPLE SILICONExplore project ↗
Fabletron · BF16
The full-precision merged model behind the quantized Fabletron exports.
MERGED WEIGHTSExplore project ↗
Qwen3.8 27B · NF4
Four-bit Qwen weights for Transformers and QLoRA fine-tuning.
BITSANDBYTESExplore project ↗
Nemotron 120B · NF4
A pre-quantized NF4 build of NVIDIA Nemotron 3 Super.
BITSANDBYTESExplore project ↗
GPT-5.6 Sol Traces
Coding, debugging, review, and defensive-security trajectories from Codex.
AGENT TRAJECTORIESExplore project ↗
GLM 5.2 Traces
GLM trajectories for instruction following, tool use, and software development.
MOONSHINER DATAExplore project ↗
Kimi K3 Traces
Kimi K3 coding, debugging, and tool-use sessions prepared for training.
MOONSHINER DATAExplore project ↗
Fable 5 · Synthetic Corrections
A companion dataset of narrowly corrected, independently re-judged traces.
CORRECTED TRACESExplore project ↗
VCF 9.x Training Data
Version-aware VMware Cloud Foundation question-and-answer training data.
INFRASTRUCTURE Q&AExplore project ↗
Kimi K3 Trace Explorer
Browse, search, and export agentic coding trajectories.
INTERACTIVE EXPLORERExplore project ↗
Fable 5 Trace Explorer
Browse, search, and export agentic coding trajectories.
INTERACTIVE EXPLORERExplore project ↗
No matching projects. Try another name or category.
Public repositories verified from GitHub and Hugging Face. Project pages contain current releases, licenses, and requirements.
THE PERSON BEHIND THE REPOS
Hi, I'm Greg.
I build things that
do the work.
My work connects infrastructure, AI agents, model training, and the tools that make them useful. This is the home for those projects—and the experiments that come next.
Read the code. Try a model. Open an issue.
Good tools get better when people use them.