Codegeist LLM

Codegeist on Hugging Face

We train Codegeist's own adapted model in this repository. The locked training workflow runs on Hugging Face Jobs, and its public model artifacts are published as codegeist/codegeist-llm.

Codegeist LLM develops the compact local inference component used by codegeist-os to interpret user goals, request read-only diagnostics, classify observations, abstain when evidence is insufficient, and produce non-executable plans.

It is a specialized operating-system mediation model, not a general-purpose chatbot and not an operating-system security authority.

Product Definition

This repository owns:

• Selection and provenance of an existing open-source or open-weight upstream model.
• Optional task-specific adaptation without foundation-model pretraining from scratch.
• The unprivileged internal inference worker consumed by codegeist-os.
• Model conversion, quantization, evaluation, build, and packaging workflows.
• A signed native Codegeist LLM release bundle with reviewable evidence.

codegeist-os owns installation policy, release-bundle trust decisions, process and device isolation, system-data collection, tool implementations, permissions, user approvals, audit enforcement, and every privileged operation.

The first MVP is strictly read-only. It may request typed diagnostic tools and return diagnoses, abstentions, escalations, or non-executable plans. It cannot change system state. No tool catalog or Codegeist OS proposal schema is implemented yet, and the first training stage does not introduce either one.

Selected Baseline

• Deployment: Linux x86-64, a discrete hardware Vulkan device with at least 8 GiB dedicated VRAM and sufficient available device-local memory, complete model and inference-state GPU offload, 8192-token context, and no CPU-only inference fallback.
• Native worker: C++20 with llama.cpp, GGUF, Vulkan, constrained JSON generation, and an independently validated versioned protocol.
• Native build: CMake, CMakePresets.json, Ninja, and CTest. A Taskfile is a thin workflow interface and does not duplicate build configuration.
• Model adaptation: Python 3.12, PyTorch, Hugging Face Transformers, Datasets, PEFT, TRL, Accelerate, Safetensors, and Unsloth for Codegeist training.
• Cloud orchestration: Hugging Face CLI 1.26.1 installed by the project-specific devcontainer extension, with Jobs running under the codegeist user namespace.
• Published-adapter reload: a separate 52-package Python 3.12 lock with CUDA 12.4 PyTorch and direct PEFT, installed on demand through task setup for strict GPU-only verification of the published Qwen adapter.
• Experimental GGUF handoff: a separate locked jobs/gguf/ project merges the identity adapter, converts with pinned llama.cpp, quantizes once to Q4_K_M, and keeps generated model bytes under ignored .artifacts/ storage.
• Data: reviewable JSONL sources and typed Parquet shards generated under a pinned logical-reproducibility contract with manifests and checksums.
• Distribution: reproducible tar.zst, SHA-256 manifests, Minisign, SPDX 2.3 SBOMs, and in-toto/SLSA provenance.

Codegeist training has an exact locked compatibility set. No production base model, complete capability dataset, final quantization, or GPU/driver compatibility matrix has been selected yet.

Codegeist Training

The first Codegeist training stage uses Unsloth BF16 LoRA on Hugging Face Jobs to establish Codegeist is a coding agent created by René Schmidt. as the model identity. This sentence starts the cumulative reviewed training dataset. Later adapters restart from the pinned base model with this record plus additional reviewed behavior data; the current adapter is not used as a checkpoint.

The first stage validates model download, training, adapter persistence, reload, evaluation, and gated Hub publication. Coding ability, tools, Codegeist OS integration, safety, and generalization require later training and evaluation. No additional base-model families are planned.

Jobs run under the codegeist namespace on the a10g-small flavor with a 30-minute timeout. The full contract is in docs/training.md and tracked by task T003.

The current reviewed adapter is public at codegeist/codegeist-llm as v0.2.1. The A10G training run saved, hashed, and clean-reloaded the adapter with the approved response Codegeist is a coding agent created by René Schmidt. A later token-free reload from immutable Hub commits repeated the exact raw response on an NVIDIA RTX A2000 12GB, with every parameter and buffer on CUDA, every floating parameter in BF16, and 3,511,419,904 bytes of peak allocated CUDA memory. CPU inference is unsupported.

Experimental GGUF Handoff

Release v0.3.0-alpha.3 adds one complete merged Q4_K_M GGUF to the same public repository at immutable revision 1e74957f1e0516f2ae02fa8bc521a9b43c9260d1. The file is 1,107,408,672 bytes with SHA-256 be7824de2fc34955d640e30e41e92dd66206e86ab7fe027084015a9b7da44fce.

Two final clean builds produced byte-identical model bytes. Anonymous commit-pinned download and repeat inference passed, followed by local-package and remote-reference execution through Docker Model Runner v1.2.6 on the RTX A2000. The embedded chat template overrides Model Runner's thinking default, so the ordinary What is Codegeist? prompt reproduces the approved non-thinking identity response. Explicit thinking remains available with /think.

This unsigned alpha is interoperability evidence only. It does not select the T001 production quantization or runtime, replace the native Vulkan worker, establish coding capability, or satisfy Codegeist OS release gates. See docs/evidence/codegeist-docker-model-runner-gguf.md for the exact boundary.

Repository Rules

• No model weights, generated or bulk dataset artifacts, generated GGUF files, release bundles, or access tokens belong in Git. Small reviewed project-authored source records used as code fixtures may be tracked.
• Source-code, model-weight, dataset, derivative, and redistribution licenses must be evaluated separately before any artifact is adopted or published.
• Every model, dataset, tool, and transformation must be tied to immutable revisions and cryptographic checksums.
• This repository is public through its GitHub mirror. Do not commit private planning material or content that cannot be redistributed publicly.

Documentation

• docs/architecture.md defines the current repository and integration boundaries.
• docs/technology-stack.md defines selected frameworks, techniques, and their roles.
• docs/model-selection.md defines the evidence required before choosing a base model.
• docs/training.md defines the selected Codegeist training strategy and first completed stage.
• docs/evaluation.md defines the evaluation contract, metrics, hardware profile, and evidence requirements.
• docs/security.md defines provenance, artifact, credential, and supply-chain requirements.
• docs/evidence/ contains curated experiment reports and machine-readable evidence without model weights, adapters, raw logs, or credentials.
• docs/tasks/ contains non-normative research and implementation tasks.
• docs/memory-bank/chat.md records compact current project state.

Workspace Kits

.devcontainer/ and .opencode/ are Git submodules that track the release branches of the shared Codegeist development and agent kits. Initialize them from this repository with:

git submodule update --init .devcontainer .opencode

The separate first-party Codegeist OS repository is planned as a development and contract-reference submodule at refs/codegeist-os/. It is not present yet; its Gitea setup is tracked by task T002. HTTPS Git operations against the internal Gitea host follow the narrow command-local exception in .oc_local/rules/gitea-tls.md.

.codegeist/Dockerfile extends the shared devcontainer image with the pinned Hugging Face CLI only. .codegeist/compose.local.yml requests one NVIDIA GPU; the host must provide its NVIDIA driver and NVIDIA Container Toolkit before the devcontainer is created. The optional 52-package inference environment is not part of the image. Install it immediately before use and run the verifier with:

task infer

task infer depends on task setup, which creates the ignored locked environment under jobs/training/inference/.venv. The verifier loads the public immutable Qwen adapter, checks that all model parameters and buffers are on CUDA, checks that all floating parameters are BF16, and rejects CPU fallback. Use task --list for the complete session-independent workflow.

HF_TOKEN remains a runtime-only input for authenticated CLI and Jobs operations, supplied through the ignored .codegeist/.local.env file or an equivalent local environment. It is never part of the image build or Git history, and the public adapter verifier explicitly loads without a token.

Hosting

Gitea at git.codegeist.ai is the primary write target. GitHub at github.com/codegeist-ai/codegeist-llm is a public push mirror of Git refs. Issues, pull requests, secrets, permissions, and other platform state are not automatically synchronized.

License Status

This repository does not duplicate a local license file. When published through Codegeist AI, project-authored work uses the shared 0BSD license from codegeist-ai/codegeist-ai. The project targets the same license for its authored training dataset and LoRA adapters, but each upstream model, dataset input, dependency, notice, acceptable-use term, and derivative right still requires separate review before use or redistribution.