Skip to content

INDEX.md

Latest commit

 

History

History
149 lines (134 loc) · 8.24 KB

INDEX.md

File metadata and controls

149 lines (134 loc) · 8.24 KB
Raw

Repository Index

Navigation map for the documentation-first Codegeist LLM workspace.

When To Read This

  • Read this before changing model-selection, evaluation, provenance, artifact, or codegeist-os integration requirements.
  • Read docs/security.md before handling upstream model or dataset artifacts.

Directory Map

  • README.md - project purpose, boundaries, hosting model, and workspace setup.
  • docs/architecture.md - normative product, component, deployment, and repository boundary.
  • docs/technology-stack.md - selected frameworks, techniques, and tool roles.
  • docs/model-selection.md - base-model and hardware-envelope decision criteria.
  • docs/training.md - Codegeist adaptation strategy and first completed Qwen3-1.7B training-stage contract.
  • docs/evaluation.md - evaluation contract, metrics, and evidence requirements.
  • docs/security.md - provenance, supply-chain, and artifact safety rules.
  • docs/evidence/ - curated experiment reports and structured evidence records; raw private artifacts remain outside Git.
  • docs/tasks/ - tracked research, architecture, and implementation tasks; T001 validates the architecture, T002 tracks the Codegeist OS reference, T003 tracks the first Codegeist training stage, and T004 tracks the experimental Docker Model Runner GGUF handoff.
  • docs/memory-bank/chat.md - compact state for future sessions.
  • jobs/training/ - locked Qwen training source, local published- adapter inference, framework probe, upstream manifest, contract tests, and Job instructions; generated adapters remain outside this directory and outside Git.
  • jobs/gguf/ - isolated locked merge, conversion, quantization, publication, and weightless contract project for the experimental Q4_K_M handoff.
  • .devcontainer/ - shared development environment on its release branch.
  • .codegeist/Dockerfile - project-specific devcontainer extension that pins the Hugging Face CLI without embedding credentials or optional inference packages.
  • .codegeist/compose.local.yml - requests one NVIDIA GPU for the project devcontainer.
  • Taskfile.yml - session-independent entrypoints for on-demand inference, contracts, evidence, GGUF build and promotion, and strict GPU reload.
  • .opencode/ - shared OpenCode agent kit on its release branch.
  • .oc_local/rules/gitea-tls.md - project-specific runtime-token authentication and command-local TLS exception for the internal Gitea host.
  • .gitmodules - shared-kit submodule sources and branch tracking.
  • refs/codegeist-os/ - planned first-party OS contract-reference submodule; absent until T002 completes Gitea repository setup.

Known Directory Indexes

  • INDEX.md - this repository-root index.

Key Workflows

  • Initialize shared kits with git submodule update --init .devcontainer .opencode from this repository.
  • Record a model-selection decision only after every required evidence category in docs/model-selection.md is addressed.
  • Use docs/tasks/T001_validate-local-os-mediation-architecture.md to validate the selected architecture and resolve remaining model and release decisions.
  • Use docs/tasks/T002_setup-codegeist-os-reference.md to add the Gitea-hosted OS repository at refs/codegeist-os/ with the narrow Gitea TLS exception and without exposing its token.
  • Use docs/tasks/T003_establish-codegeist-training.md and docs/training.md for the first Codegeist training stage and later-stage requirements.
  • Use docs/tasks/T004_publish-docker-model-runner-gguf.md for the separated local GPU build and promotion of the experimental merged Q4_K_M GGUF; do not treat it as T001 Vulkan or production-release evidence.
  • Read docs/evidence/codegeist-training-qwen3-1.7b.md for the paid run, publication, cost, artifact hashes, anonymous GPU evidence, and limits.
  • Read docs/evidence/codegeist-training-overview.md for the generated visual dashboard, loss curve, provenance chain, and interpretation boundary.
  • Read docs/evidence/codegeist-docker-model-runner-gguf.md for the immutable GGUF identity, reproducible builds, anonymous download, Docker GPU result, and Qwen thinking-mode control.
  • Run task test for all four lock checks and 52 weightless training and GGUF contracts; use jobs/training/README.md for the paid Jobs gate and jobs/gguf/README.md for the separated GGUF workflow.
  • Load the reviewed Qwen adapter from codegeist/codegeist-llm, pinning adapter commit a9504a0ee1150ea05f88ff725758404fcb604a32 and the base revision recorded in jobs/training/README.md.
  • Rebuild the devcontainer on a host with the NVIDIA driver and NVIDIA Container Toolkit. Run task infer to install the optional locked environment on demand and perform the token-free public-adapter reload. Verify hf version, hf auth whoami, and hf jobs hardware without printing HF_TOKEN before authenticated Jobs work.
  • Keep model and dataset artifacts outside Git and commit only reviewable manifests, checksums, and documentation when an artifact workflow exists.

Search Hints

  • hardware envelope - measurements required for consumer-hardware targets.
  • provenance - source and transformation evidence requirements.
  • deferred - open decisions and later release gates.
  • codegeist-os - model-to-operating-system integration boundary.
  • local OS mediation, typed tool request, read-only MVP, inference worker: product boundaries defined in docs/architecture.md.
  • Vulkan, GGUF, downloadable native distribution, Qwen3 - deployment baseline and single-candidate evaluation terms.
  • Unsloth, Hugging Face Jobs, a10g-small, Codegeist Training, Codegeist is a coding agent created by René Schmidt. - first-stage training terms.
  • hf==1.26.1, HF_TOKEN, codegeist, .codegeist/Dockerfile, 70d244cc86ccca08cf5af4e1e306ecf908b1ad5e - first Jobs preflight and Qwen initial training inputs.
  • Taskfile.yml, task setup, task infer, inference/.venv - on-demand inference installation and strict CUDA BF16 reload outside the image.
  • render_training_evidence.py, codegeist-training-dashboard.svg, codegeist-training-provenance.mmd, codegeist-training-provenance.svg - generated visual evidence overview and Mermaid provenance.
  • 6a76c9983e1f34a7e32be58c, 4cc89bd25712ff4f, a9504a0ee1150ea, c039e9013856f, v0.2.1, af0092e72bd347d5 - initial training, publication, and anonymous local GPU evidence.
  • T004, Q4_K_M, Docker Model Runner, docker model package, v0.3.0-alpha.3, 1e74957f1e05, be7824de2fc3, /think, b10333, 08659901c43b - experimental merged GGUF build, publication, anonymous integrity check, and GPU local-package handoff terms.
  • refs/codegeist-os, GITEA_TOKEN, credential.helper, gitea-tls.md, sslVerify=false - planned first-party reference-submodule setup, ephemeral authentication, and narrow transport exception.
  • CMake, CMakePresets.json, Taskfile, full GPU offload, 8 GiB VRAM, 8192-token context, PyTorch, PEFT, TRL, Minisign, SPDX - selected stack and first deployment-profile terms in docs/technology-stack.md.

Update Triggers

  • Update this index when major documents or top-level directories change.
  • Update task entries when an architecture task creates, moves, or retires durable documentation.
  • Update docs/memory-bank/chat.md when a model, runtime, hardware envelope, or artifact contract is selected.

Agent Notes

  • Do not add weights, generated or bulk dataset artifacts, generated model artifacts, or credentials. Small reviewed project-authored source records may be tracked as code fixtures.
  • Public project-authored content uses the shared 0BSD license from codegeist-ai/codegeist-ai; this repository does not duplicate its license file. Third-party models, datasets, dependencies, and derivative rights still require separate review.
  • Do not represent the first training stage as evidence of coding ability, tool use, generalization, local deployment, or production model quality.
  • Before any HTTPS Git or API operation against git.codegeist.ai, read .oc_local/rules/gitea-tls.md. Pass runtime GITEA_TOKEN only through its ephemeral credential path and use only its exact host-scoped, command-local TLS exception; never persist or broaden either mechanism.