DLM — Document Language Model
Turn a text file into a fine-tuned LLM. Edit the document, retrain, share.
A .dlm file is YAML frontmatter + markdown body with ::instruction:: and ::preference:: sections. Train a LoRA adapter on your document, prompt it, export to Ollama. This extension makes .dlm authoring a first-class editor experience.
Features
Syntax-aware editing
YAML frontmatter, markdown body, and section fences (::instruction::, ::preference::, ::image::, ::audio::) get distinct, theme-aware highlighting. Q/A headers inside instruction blocks are styled.
Smart completions
Autocomplete for the 26-entry base-model registry, adapter types, quantization levels, and section fences — all driven by the live schema.
Hover insight
Hover any base-model key to see params, VRAM footprint, context length, modality, and license. Hover a section fence for the section ID and quick stats.
Live diagnostics
Schema errors, unknown base models, and Doctor warnings (VRAM headroom, unsafe combinations) surface inline as you type — no save required.
Side panel — frictionless authoring
A dedicated DLM activity-bar view with everything you need to compose a .dlm:
- Quick Insert — Instruction, Preference, Image, and Audio buttons drop snippet templates with tab stops at the cursor
- Source Directory Manager — native folder picker; relative paths computed and inserted into frontmatter via the LSP
- Base Model Browser — searchable QuickPick over the full registry; click to set
base_model
- Template Gallery — 8 bundled templates with one-click bootstrap
- Document Overview — section counts, store status, adapter version
- Training Controls — Run / Stop in an integrated terminal
Transparent store creation
Open a .dlm file. The content-addressed store at ~/.dlm/store/<dlm_id>/ is created on the spot — no manual dlm init needed.
Requirements
Python 3.11+ with the document-language-model package:
pip install document-language-model
dlm-lsp language server:
pip install dlm-lsp
(Optional) For training, choose a hardware-appropriate base. The Doctor panel will tell you what fits.
Quick start
- Install the extension.
- Create a new file with a
.dlm extension. The side panel becomes available immediately.
- Click Add Instruction in the side panel and fill in the Q / A.
- Open the Base Model Browser and pick a small base (e.g.
qwen2.5-0.5b) for a fast first run.
- Run
DLM: Train Current Document from the command palette (⇧⌘P / Ctrl+Shift+P).
- Run
DLM: Prompt to chat with your trained adapter.
- Run
DLM: Export to package an Ollama-ready Modelfile + GGUF.
Commands
| Command |
What it does |
DLM: Train Current Document |
Fine-tune the LoRA adapter on the open .dlm |
DLM: Export |
Export base.gguf + adapter.gguf + Modelfile |
DLM: Synth Instructions |
Generate synthetic Q/A pairs from your sources |
DLM: Show Run History |
Browse training runs for the current document |
DLM: Open Store Directory |
Reveal ~/.dlm/store/<id> in the file manager |
DLM: Insert Instruction Section |
Drop an instruction block at the cursor |
DLM: Insert Preference Section |
Drop a preference block at the cursor |
Settings
| Setting |
Default |
Description |
dlm.command |
uv run dlm |
Command to invoke the dlm CLI (use dlm for system venv) |
dlm.home |
"" |
Override the ~/.dlm store root |
dlm.defaultBase |
"" |
Default base_model for new documents |
dlm.watchOnSave |
false |
Auto-retrain on save |
dlm.lspPath |
dlm-lsp |
Path to the dlm-lsp binary |
If dlm is on your PATH (e.g. installed system-wide), set dlm.command to dlm. The default uv run dlm works inside a uv-managed project.
Multi-editor support
The language server is editor-agnostic. Cookbook entries for Zed, Helix, and Neovim live in the main repository.
Links
License
MIT
