Ollama ColBERT Local Artifacts
This repository tracks local artifact notes for Ollama ColBERT and related BitNet setup work. It is documentation-only at this stage: no third-party model weights, GGUF files, or derived binaries are uploaded here.
Status
Experimental public handoff repository for local validation and artifact planning.
GitHub remains the source of truth for code, review, branches, and pull requests. Hugging Face is used here for model-facing notes, artifact manifests, and future model-card context.
Source Links
| Area | Link |
|---|---|
| Ollama ColBERT upstream PR | https://github.com/ollama/ollama/pull/16195 |
| Ollama handoff branch | https://github.com/edithatogo/ollama/tree/conductor/colbert-handoff |
| BitNet local wrapper branch | https://github.com/edithatogo/BitNet/tree/local-bitnet-runner-wrapper |
| BitNet generated-header ignore PR | https://github.com/microsoft/BitNet/pull/563 |
| Current publication status | status.md |
| Validation evidence | validation.md |
What This Covers
- Local Ollama support work for LFM2/ColBERT-style embedding checkpoints.
- Handoff notes for validating ColBERT model loading and embedding behavior.
- Local BitNet setup notes for a Microsoft BitNet GGUF through the official BitNet runner.
- A conservative artifact manifest that separates safe documentation from blocked model-weight uploads.
Local Ollama Validation
The Ollama work is tracked in GitHub. See validation.md for the current post-rebase validation evidence. The key validation commands used for the code branch were:
go test -count=1 ./cmd ./convert ./parser
find . -name '*.go' \
-not -path './app/*' \
-not -path './integration/*' \
-not -path './.git/*' \
-exec dirname {} \; | sort -u | sed 's#^\./#./#' | xargs go test -count=1
As of the June 14 refresh, the upstream PR is a focused conversion/docs change. Older local runtime and GGUF metadata patching experiments are documented in the Conductor handoff, but they are not part of the current PR diff.
Local BitNet Notes
The BitNet wrapper branch is a personal-fork convenience layer for local use. It is not proposed upstream as-is because it assumes local paths, a repo-local virtual environment, and Unix shell conventions.
Example local invocation from the fork branch:
bin/bitnet "Say ok" -n 1 -c 128 -temp 0
The upstream-suitable generated-header ignore change was split into a separate Microsoft BitNet pull request.
Local Apple MLX Notes
Apple MLX is available locally through a dedicated virtualenv at /Volumes/PortableSSD/GitHub/.venvs/mlx-lm. The verified packages are mlx==0.31.2 and mlx-lm==0.31.3, with MLX selecting Device(gpu, 0) on this machine.
Example local CLI check:
/Volumes/PortableSSD/GitHub/.venvs/mlx-lm/bin/python -m mlx_lm --help
Artifact Policy
This repository does not redistribute:
- Microsoft BitNet model weights.
- LiquidAI model weights.
- Generated GGUF files derived from third-party model weights.
Those artifacts should only be uploaded after source license, redistribution permissions, and provenance are reviewed and documented in artifact-manifest.md.
Source Model Metadata Checked
The source model metadata was checked on 2026-05-24 before publication:
| Source model | Hugging Face metadata | Publication decision here |
|---|---|---|
LiquidAI/LFM2-ColBERT-350M |
license:other, public, ungated, sha 0c31032e995fe698f3ddd74f0ddb566cfd3d4d5a |
Reference only; no weights or derived GGUF uploaded. |
microsoft/bitnet-b1.58-2B-4T |
license:mit, public, ungated, sha 04c3b9ad9361b824064a1f25ea60a8be9599b127 |
Reference only; no weights or derived GGUF uploaded. |
Recommended Use
Use this repository as a model-facing index for the GitHub work:
- Follow the GitHub PRs for source changes.
- Use this Hugging Face repo for artifact status and usage notes.
- Add GGUF/model files only after license and provenance checks are complete.