documentlanguagemodel Public

Watch 0 Fork 0 Star 0

markdown · 3268 bytes Raw Blame History

Template gallery

The fastest path from "I want to try dlm" to a trained adapter is dlm init --template <name>. Each template is a curated .dlm file with a sensible base model, body content to train on, and sample prompts to exercise after training.

Listing the gallery

$ dlm templates list
changelog                 Changelog entry generator              (smollm2-360m)
coding-tutor              Coding tutor (Python, curated)         (qwen2.5-coder-1.5b)
domain-kb                 Domain knowledge base                  (qwen2.5-3b)
meeting-notes-summarizer  Meeting notes → decision log           (qwen2.5-3b)
personal-assistant        Personal assistant                     (qwen2.5-1.5b)
regex-buddy               Regex explainer                        (qwen2.5-coder-1.5b)
shell-one-liner           Shell one-liner helper                 (qwen2.5-coder-1.5b)
writing-partner           Writing partner (stylistic continuation) (llama-3.2-3b)

Pass --json for machine-readable output (every field of the template's meta.yaml including domain_tags, expected_steps, expected_duration, and sample_prompts).

Creating a new `.dlm` from a template

$ dlm init mytutor.dlm --template coding-tutor
init: wrote mytutor.dlm from template coding-tutor (Coding tutor
(Python, curated)) — base qwen2.5-coder-1.5b.

$ dlm train mytutor.dlm
$ dlm prompt mytutor.dlm "What are Python decorators?"

The template's recommended_base is adopted automatically. If you pass --base alongside --template, the template wins (and you get a note); the bundled body was authored against its recommended base, so swapping bases mid-body is an advanced move — edit the frontmatter yourself after init if that's what you want.

What's in a template

Each template is a pair of files:

coding-tutor.dlm        # the `.dlm` body: frontmatter + sections
coding-tutor.meta.yaml  # metadata — name, title, tags, recommended_base, summary

The meta.yaml schema:

name: coding-tutor               # filename stem, required to match
title: Coding tutor (Python, curated)
domain_tags: [code, python, tutor]
recommended_base: qwen2.5-coder-1.5b
expected_steps: 800              # rough step count at defaults
expected_duration:               # hardware-tier → wall-clock estimate
  cuda-sm80+: "~5 min"
  mps: "~15 min"
  cpu: "~2 hr"
summary: |
  A compact Python-focused Q&A tutor.
sample_prompts:
  - "What are Python decorators?"

Invalid metadata (unknown keys, missing required fields, sidecar absent) drops the template from the listing with a logged warning — never silently served.

Where the gallery lives

The bundled gallery ships inside the dlm package at src/dlm/templates/gallery/. importlib.resources.files resolves the path regardless of whether you're running from a source checkout or a pip install-ed wheel.

Remote gallery — deferred

The sprint spec calls for dlm templates list --refresh to pull from github.com/<org>/dlm-templates with signed-tag verification against a pinned public key. That upstream repo and signing key are pending; until they land, --refresh emits a clear warning and falls back to the bundled gallery. All usage here works offline today.

View source

  
        1
        # Template gallery
      
        2
        
        3
        The fastest path from "I want to try `dlm`" to a trained adapter is
      
        4
        `dlm init --template <name>`. Each template is a curated `.dlm` file
      
        5
        with a sensible base model, body content to train on, and sample
      
        6
        prompts to exercise after training.
      
        7
        
        8
        ## Listing the gallery
      
        9
        
        10
        ```bash
      
        11
        $ dlm templates list
      
        12
        changelog                 Changelog entry generator              (smollm2-360m)
      
        13
        coding-tutor              Coding tutor (Python, curated)         (qwen2.5-coder-1.5b)
      
        14
        domain-kb                 Domain knowledge base                  (qwen2.5-3b)
      
        15
        meeting-notes-summarizer  Meeting notes → decision log           (qwen2.5-3b)
      
        16
        personal-assistant        Personal assistant                     (qwen2.5-1.5b)
      
        17
        regex-buddy               Regex explainer                        (qwen2.5-coder-1.5b)
      
        18
        shell-one-liner           Shell one-liner helper                 (qwen2.5-coder-1.5b)
      
        19
        writing-partner           Writing partner (stylistic continuation) (llama-3.2-3b)
      
        20
        ```
      
        21
        
        22
        Pass `--json` for machine-readable output (every field of the template's
      
        23
        `meta.yaml` including `domain_tags`, `expected_steps`,
      
        24
        `expected_duration`, and `sample_prompts`).
      
        25
        
        26
        ## Creating a new `.dlm` from a template
      
        27
        
        28
        ```bash
      
        29
        $ dlm init mytutor.dlm --template coding-tutor
      
        30
        init: wrote mytutor.dlm from template coding-tutor (Coding tutor
      
        31
        (Python, curated)) — base qwen2.5-coder-1.5b.
      
        32
        
        33
        $ dlm train mytutor.dlm
      
        34
        $ dlm prompt mytutor.dlm "What are Python decorators?"
      
        35
        ```
      
        36
        
        37
        The template's `recommended_base` is adopted automatically. If you pass
      
        38
        `--base` alongside `--template`, the template wins (and you get a
      
        39
        note); the bundled body was authored against its recommended base, so
      
        40
        swapping bases mid-body is an advanced move — edit the frontmatter
      
        41
        yourself after `init` if that's what you want.
      
        42
        
        43
        ## What's in a template
      
        44
        
        45
        Each template is a pair of files:
      
        46
        
        47
        ```
      
        48
        coding-tutor.dlm        # the `.dlm` body: frontmatter + sections
      
        49
        coding-tutor.meta.yaml  # metadata — name, title, tags, recommended_base, summary
      
        50
        ```
      
        51
        
        52
        The `meta.yaml` schema:
      
        53
        
        54
        ```yaml
      
        55
        name: coding-tutor               # filename stem, required to match
      
        56
        title: Coding tutor (Python, curated)
      
        57
        domain_tags: [code, python, tutor]
      
        58
        recommended_base: qwen2.5-coder-1.5b
      
        59
        expected_steps: 800              # rough step count at defaults
      
        60
        expected_duration:               # hardware-tier → wall-clock estimate
      
        61
          cuda-sm80+: "~5 min"
      
        62
          mps: "~15 min"
      
        63
          cpu: "~2 hr"
      
        64
        summary: |
      
        65
          A compact Python-focused Q&A tutor.
      
        66
        sample_prompts:
      
        67
          - "What are Python decorators?"
      
        68
        ```
      
        69
        
        70
        Invalid metadata (unknown keys, missing required fields, sidecar
      
        71
        absent) drops the template from the listing with a logged warning —
      
        72
        never silently served.
      
        73
        
        74
        ## Where the gallery lives
      
        75
        
        76
        The bundled gallery ships inside the `dlm` package at
      
        77
        `src/dlm/templates/gallery/`. `importlib.resources.files` resolves the
      
        78
        path regardless of whether you're running from a source checkout or a
      
        79
        `pip install`-ed wheel.
      
        80
        
        81
        ## Remote gallery — deferred
      
        82
        
        83
        The sprint spec calls for `dlm templates list --refresh` to pull from
      
        84
        `github.com/<org>/dlm-templates` with signed-tag verification against a
      
        85
        pinned public key. That upstream repo and signing key are pending;
      
        86
        until they land, `--refresh` emits a clear warning and falls back to
      
        87
        the bundled gallery. All usage here works offline today.

1	# Template gallery
2
3	The fastest path from "I want to try `dlm`" to a trained adapter is
4	`dlm init --template <name>`. Each template is a curated `.dlm` file
5	with a sensible base model, body content to train on, and sample
6	prompts to exercise after training.
7
8	## Listing the gallery
9
10	```bash
11	$ dlm templates list
12	changelog Changelog entry generator (smollm2-360m)
13	coding-tutor Coding tutor (Python, curated) (qwen2.5-coder-1.5b)
14	domain-kb Domain knowledge base (qwen2.5-3b)
15	meeting-notes-summarizer Meeting notes → decision log (qwen2.5-3b)
16	personal-assistant Personal assistant (qwen2.5-1.5b)
17	regex-buddy Regex explainer (qwen2.5-coder-1.5b)
18	shell-one-liner Shell one-liner helper (qwen2.5-coder-1.5b)
19	writing-partner Writing partner (stylistic continuation) (llama-3.2-3b)
20	```
21
22	Pass `--json` for machine-readable output (every field of the template's
23	`meta.yaml` including `domain_tags`, `expected_steps`,
24	`expected_duration`, and `sample_prompts`).
25
26	## Creating a new `.dlm` from a template
27
28	```bash
29	$ dlm init mytutor.dlm --template coding-tutor
30	init: wrote mytutor.dlm from template coding-tutor (Coding tutor
31	(Python, curated)) — base qwen2.5-coder-1.5b.
32
33	$ dlm train mytutor.dlm
34	$ dlm prompt mytutor.dlm "What are Python decorators?"
35	```
36
37	The template's `recommended_base` is adopted automatically. If you pass
38	`--base` alongside `--template`, the template wins (and you get a
39	note); the bundled body was authored against its recommended base, so
40	swapping bases mid-body is an advanced move — edit the frontmatter
41	yourself after `init` if that's what you want.
42
43	## What's in a template
44
45	Each template is a pair of files:
46
47	```
48	coding-tutor.dlm # the `.dlm` body: frontmatter + sections
49	coding-tutor.meta.yaml # metadata — name, title, tags, recommended_base, summary
50	```
51
52	The `meta.yaml` schema:
53
54	```yaml
55	name: coding-tutor # filename stem, required to match
56	title: Coding tutor (Python, curated)
57	domain_tags: [code, python, tutor]
58	recommended_base: qwen2.5-coder-1.5b
59	expected_steps: 800 # rough step count at defaults
60	expected_duration: # hardware-tier → wall-clock estimate
61	cuda-sm80+: "~5 min"
62	mps: "~15 min"
63	cpu: "~2 hr"
64	summary: \|
65	A compact Python-focused Q&A tutor.
66	sample_prompts:
67	- "What are Python decorators?"
68	```
69
70	Invalid metadata (unknown keys, missing required fields, sidecar
71	absent) drops the template from the listing with a logged warning —
72	never silently served.
73
74	## Where the gallery lives
75
76	The bundled gallery ships inside the `dlm` package at
77	`src/dlm/templates/gallery/`. `importlib.resources.files` resolves the
78	path regardless of whether you're running from a source checkout or a
79	`pip install`-ed wheel.
80
81	## Remote gallery — deferred
82
83	The sprint spec calls for `dlm templates list --refresh` to pull from
84	`github.com/<org>/dlm-templates` with signed-tag verification against a
85	pinned public key. That upstream repo and signing key are pending;
86	until they land, `--refresh` emits a clear warning and falls back to
87	the bundled gallery. All usage here works offline today.