Skip to content
Merged
Show file tree
Hide file tree
Changes from all commits
Commits
File filter

Filter by extension

Filter by extension

Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
1 change: 0 additions & 1 deletion README.md
Original file line number Diff line number Diff line change
Expand Up @@ -49,7 +49,6 @@ See the official [plugin documentation](https://www.appdevforall.org/codeonthego
| [`plugins/Voice-Alerts/`](plugins/Voice-Alerts/) | Plays a sound when a build finishes or fails. |

> `cotg-ndk` and `pebble-custom-function-template-installer` are not listed. They are still at the repository root, beside `libs/` and `tools/`, until their own work lands.
> The `AI-*` addons above build from `plugins/` like every other example but are still held out of the published gallery; `tools/addons/skip.txt` states the reason for each.

## Templates

Expand Down
10 changes: 5 additions & 5 deletions plugins/Get-AI-Models/README.md
Original file line number Diff line number Diff line change
Expand Up @@ -135,15 +135,15 @@ and the GGUF re-upload; loads and generates coherent output in our AI plugin; pa
tool-use eval at ≥ 80%; valid chat template; single file from a reputable uploader with the SHA-256
pinned to a specific revision; context window ≥ 16k tokens. Quantization is *not* restricted.

**The current catalog does not satisfy all of them.** The maintainer selected these six models
**The current catalog does not satisfy all of them.** The maintainer selected these models
explicitly, waiving the instruction-tuned, open-licence and 16k-context gates for specific entries —
so `CatalogLoader` no longer enforces the licence or context gates, and the unit tests no longer
assert them. `docs/CURATION.md` records exactly which gate each entry breaks.

The catalog lists 7 files: Qwen2.5 0.5B (base), Qwen3 0.6B, Qwen3.5 0.8B, Llama 3.2 1B Instruct,
SmolLM2 360M Instruct, H2O Danube3 500M (base), all under 1 GB, and Qwen2.5-Coder 1.5B Instruct
(1.04 GiB), the only code-trained model. Six are Apache-2.0;
Llama 3.2 1B is under the Llama 3.2 Community Licence.
The catalog lists 10 files: Qwen2.5 0.5B (base), Qwen3 0.6B, Qwen3.5 0.8B, Llama 3.2 1B Instruct,
SmolLM2 360M Instruct, H2O Danube3 500M (base), OLMo 2 1B Instruct, Granite 4.0 1B and
SmolLM2 1.7B Instruct, all under 1 GiB, and Qwen2.5-Coder 1.5B Instruct (1.04 GiB), the only
code-trained model. Nine are Apache-2.0; Llama 3.2 1B is under the Llama 3.2 Community Licence.

**The two behavioural gates are not yet proven.** Per the ticket, the admission harness that runs a
candidate through the AI plugin is downstream work, so the shipped catalog is the set of candidates
Expand Down
49 changes: 38 additions & 11 deletions plugins/Get-AI-Models/docs/CURATION.md
Original file line number Diff line number Diff line change
Expand Up @@ -51,17 +51,21 @@ Metadata came from the Hugging Face API, not from model cards or memory:
`sha256` comes from — no multi-gigabyte download was needed to obtain it.
- `GET /{base}/resolve/main/config.json` — `max_position_embeddings` (gate 7)
- `GET /{base}/resolve/main/tokenizer_config.json` (or `chat_template.jinja`) — gate 5
- `HEAD` on each pinned `resolve/{sha}/{file}` URL — confirmed HTTP 200 with `Content-Length`
equal to the catalogued `sizeBytes` and `X-Linked-Etag` equal to the catalogued `sha256`.
- `HEAD` on each pinned `resolve/{sha}/{file}` URL — confirmed the HTTP 302 to the CDN carries
`X-Linked-Size` equal to the catalogued `sizeBytes` and `X-Linked-Etag` equal to the catalogued
`sha256`. (`Content-Length` on that 302 is the redirect body, not the file.)

All 12 entries passed that final HEAD cross-check, re-run after the most recent edit.
All 10 entries passed that final HEAD cross-check, re-run on 2026-11-18 after the most recent edit.

## Catalog entries (maintainer-selected)

**This catalog is a maintainer override, not a gate-derived list.** The first six models below were
named explicitly by the ticket owner, replacing the eleven gate-passing entries that shipped
earlier; the seventh was requested later in ADFA-6321. Three of the six break at least one hard gate; those gates are therefore **no longer enforced** in
`CatalogLoader` (licence, 16k context) nor asserted in `CatalogLoaderTest`.
**This catalog is a maintainer override, not a gate-derived list.** It has ten entries. The first
six in the table below were named explicitly by the ticket owner, replacing the eleven gate-passing
entries that shipped earlier; `qwen2-5-coder-1-5b-instruct-q4_k_m` was requested later in
ADFA-6321, and the last three in ADFA-7213. Six of the ten break at least one hard gate — four of
the owner's first six, plus SmolLM2 1.7B and OLMo 2 1B — as the *Gate waived* column shows. The
licence and 16k-context gates are therefore **no longer enforced** in `CatalogLoader` nor asserted
in `CatalogLoaderTest`.

| Entry id | Base model | GGUF uploader | Ctx | Licence | Gate waived |
|---|---|---|---|---|---|
Expand All @@ -72,13 +76,36 @@ earlier; the seventh was requested later in ADFA-6321. Three of the six break at
| `smollm2-360m-instruct-q8_0` | `HuggingFaceTB/SmolLM2-360M-Instruct` | Hugging Face TB | 8192 | apache-2.0 | **7** (8192 context) |
| `h2o-danube3-500m-base-q4_k_m` | `h2oai/h2o-danube3-500m-base` | mradermacher | 8192 | apache-2.0 | **1** (base model), **7** (8192 context) |
| `qwen2-5-coder-1-5b-instruct-q4_k_m` | `Qwen/Qwen2.5-Coder-1.5B-Instruct` | Qwen (Alibaba Cloud) | 32768 | apache-2.0 | none |
| `olmo-2-0425-1b-instruct-q4_k_m` | `allenai/OLMo-2-0425-1B-Instruct` | Ai2 (allenai) | 4096 | apache-2.0 (base only) | **2** (GGUF repo declares no licence), **7** (4096 context) |
| `granite-4.0-1b-q4_k_m` | `ibm-granite/granite-4.0-1b` | IBM Granite | 131072 | apache-2.0 | none |
| `smollm2-1.7b-instruct-q4_k_m` | `HuggingFaceTB/SmolLM2-1.7B-Instruct` | Hugging Face TB | 8192 | apache-2.0 | **7** (8192 context) |

`qwen2-5-coder-1-5b-instruct-q4_k_m` was added by ADFA-6321: it is the only code-trained model in
the catalog. Its gates 1, 2, 5, 6 and 7 were re-checked against the API on 2026-10-02, and
`minRamBytes` is 4 GiB because it is the largest file here (1.04 GiB), one step above the 3 GiB of
the ~0.8 GB entries.

Gates 3 and 4 remain unproven for all seven, as before. Gate 5 (chat template) does not apply to the
The last three were added by ADFA-7213 to give fully open, non-Qwen choices: SmolLM2 1.7B Instruct
as the recommended model, OLMo 2 1B and a small Granite for the grant's model tests. Gates 1, 2, 5, 6
and 7 were checked against the API on 2026-11-18. Notes:

- **SmolLM2 1.7B Instruct** was listed below as rejected on gate 7; the owner now asks for it, so it
joins the 360M on the same context waiver. Official Q4_K_M GGUF.
- **OLMo 2 1B Instruct** uses Ai2's own GGUF repository, whose `cardData` carries only `base_model`,
no `license`. Gate 2 is waived for it on the strength of the base repo's Apache-2.0 — the same
shape that keeps `bartowski/microsoft_Phi-4-mini-instruct-GGUF` rejected, accepted here because
the uploader is the model's author. Published parameter count is 1.48B, so the row says 1.5B rather than the name's "1B".
- **Granite 4.0 1B** is the dense variant: `config.json` says `granitemoehybrid`, but every one of the
40 `layer_types` is `attention` and the GGUF architecture is `granite`. The hybrid-Mamba
`granite-4.0-h-1b` (0.90 GB) was passed over because llama.cpp's `granitehybrid` support is newer.
Its chat template is in `chat_template.jinja`, not `tokenizer_config.json`, and is embedded in the
GGUF. Published parameter count is 1.63B, so the row says 1.6B rather than the name's "1B". It is
the only one of the three that clears every statically checkable gate.

All three take `minRamBytes` 4 GiB, as Qwen2.5-Coder does: they are the same ~1 GB, ~1.5B-parameter
class.

Gates 3 and 4 remain unproven for all ten, as before. Gate 5 (chat template) does not apply to the
two base models - they have none, which is part of why they are not usable as assistants. Gate 6
holds for every entry: single file, sha256 pinned to a commit revision, HEAD-verified.

Expand Down Expand Up @@ -106,7 +133,6 @@ Each row's evidence was read from the same API, not from recollection.
| `Qwen/Qwen2.5-3B-Instruct` | 2 | `license: other`, `license_name: qwen-research` — unlike its Apache-2.0 1.5B and 7B siblings. |
| `tiiuae/Falcon3-3B-Instruct` | 2 | `license: other`, `license_name: falcon-llm-license`; carries an acceptable-use policy. |
| `bartowski/microsoft_Phi-4-mini-instruct-GGUF` | 2 | The base model is MIT, but the re-upload repository declares no licence at all. Gate 2 requires the licence on **both**, so it is out until the re-upload states one. |
| `HuggingFaceTB/SmolLM2-1.7B-Instruct` | 7 | `max_position_embeddings: 8192`. |
| `allenai/OLMo-2-1124-7B-Instruct` | 7 | `max_position_embeddings: 4096`. |
| `openai/gpt-oss-20b` | — | Apache-2.0, ungated, 131072 context: it clears every statically checkable gate. Held back on size — roughly 12 GB at MXFP4 is not usable on the phones this plugin targets — pending the deferred runtime-budget gate. Not a licence rejection. |
| `google/gemma-3-270m` | 2, plus unavailable | `license: gemma`, `gated: manual`. **Requested by the owner but not addable:** the repo ships no GGUF, and the only GGUF conversions are of the `-it` variant, which is a different model. Substituting it was not authorised. |
Expand All @@ -119,8 +145,9 @@ Each row's evidence was read from the same API, not from recollection.
want from `/api/models/{repo}/tree/{sha}?recursive=true`.
3. Build the URL as `https://huggingface.co/{repo}/resolve/{sha}/{file}` — a commit sha, never
`main`, or a re-upload would break the pinned checksum.
4. `HEAD` that URL and confirm `Content-Length` matches `size` and `X-Linked-Etag` matches
`lfs.oid`.
4. `HEAD` that URL (without following the redirect) and confirm `X-Linked-Size` matches `size`
and `X-Linked-Etag` matches `lfs.oid`. `Content-Length` on that 302 is the redirect body, not
the file.
5. Write the strengths-and-weaknesses paragraph for **that quantization**, not for the model family,
and say what it is bad at.
6. Add the row here, then run `../gradlew test` — `CatalogLoaderTest` re-validates the whole asset.
29 changes: 28 additions & 1 deletion plugins/Get-AI-Models/get-ai-models.html
Original file line number Diff line number Diff line change
Expand Up @@ -68,7 +68,7 @@ <h3>The catalog</h3>
one are genuinely different products. Each entry carries a name, quantization, parameter count,
exact on-disk byte size, a strengths-and-weaknesses paragraph, the direct <code>.gguf</code> URL, a
SHA-256, a minimum-RAM figure, and the publisher.</p>
<p>What ships in version 1.0.0:</p>
<p>What the catalog ships:</p>
<table>
<tr>
<th>Model</th><th>Quant</th><th>Params</th><th>On disk</th><th>Context</th>
Expand Down Expand Up @@ -128,6 +128,33 @@ <h3>The catalog</h3>
<td>apache-2.0</td>
<td>mradermacher</td>
</tr>
<tr>
<td>OLMo 2 1B Instruct</td>
<td><code>Q4_K_M</code></td>
<td>1.5B</td>
<td>0.87 GB</td>
<td>4k</td>
<td>apache-2.0</td>
Comment thread
hal-eisen-adfa marked this conversation as resolved.
Resolved
<td>Ai2</td>
</tr>
<tr>
<td>Granite 4.0 1B</td>
<td><code>Q4_K_M</code></td>
<td>1.6B</td>
<td>0.95 GB</td>
<td>128k</td>
<td>apache-2.0</td>
<td>IBM Granite</td>
</tr>
<tr>
<td>SmolLM2 1.7B Instruct</td>
<td><code>Q4_K_M</code></td>
<td>1.7B</td>
<td>0.98 GB</td>
<td>8k</td>
<td>apache-2.0</td>
<td>Hugging Face TB</td>
</tr>
<tr>
<td>Qwen2.5-Coder 1.5B Instruct</td>
<td><code>Q4_K_M</code></td>
Expand Down
51 changes: 51 additions & 0 deletions plugins/Get-AI-Models/src/main/assets/catalog/models.json
Original file line number Diff line number Diff line change
Expand Up @@ -103,6 +103,57 @@
"description": "H2O's small pretrained model, Apache-2.0. A base model, NOT instruction-tuned, so it continues text instead of answering, and it has an 8k context window. Strengths: very small, permissively licensed, trained on a well-described corpus. Weaknesses: unusable as an assistant as shipped, no chat template, no tool use, and below the 16k context floor. Listed at the ticket owner's request; it meets neither the instruction-tuned nor the context gate.",
"behaviouralGatesVerified": false
},
{
"id": "olmo-2-0425-1b-instruct-q4_k_m",
"name": "OLMo 2 1B Instruct",
"quantization": "Q4_K_M",
"parameters": "1.5B",
"fileName": "OLMo-2-0425-1B-Instruct-Q4_K_M.gguf",
"sizeBytes": 935515296,
"sha256": "abd8187934a438fbf7cfff0a1de5b9d2793ce913f158794df1951dcba6c93cc6",
"url": "https://huggingface.co/allenai/OLMo-2-0425-1B-Instruct-GGUF/resolve/62f8c199538474c3e33ed5d7e0580abd66686a27/OLMo-2-0425-1B-Instruct-Q4_K_M.gguf",
"minRamBytes": 4294967296,
"publisher": "Ai2 (Allen Institute for AI) - official GGUF release",
"contextTokens": 4096,
"license": "apache-2.0",
Comment thread
hal-eisen-adfa marked this conversation as resolved.
Resolved
"baseModel": "allenai/OLMo-2-0425-1B-Instruct",
"description": "Ai2's smallest instruct model and the most open one here: weights, training data, code and training recipe are all public, so it is the one to choose when the question is how a model was made. Strengths: Apache-2.0, an official GGUF from Ai2 itself, and a chat template. Weaknesses: a 4k context window - the smallest in this catalog - so only short files and short conversations fit; at about 1.5B parameters with Q4_K_M rounding it is weak at code and multi-step reasoning, and its model card claims no tool-call support. Listed at the ticket owner's request despite the context gate; the GGUF repository also declares no licence of its own.",
"behaviouralGatesVerified": false
},
{
"id": "granite-4.0-1b-q4_k_m",
"name": "Granite 4.0 1B",
"quantization": "Q4_K_M",
"parameters": "1.6B",
"fileName": "granite-4.0-1b-Q4_K_M.gguf",
"sizeBytes": 1023645440,
"sha256": "22ec0f9cc99a90185312de3c882c84e7bd6789bdd050389844380a01a831d7f1",
"url": "https://huggingface.co/ibm-granite/granite-4.0-1b-GGUF/resolve/b27c2fe3f211b7f44e80fa620177aea371099aaa/granite-4.0-1b-Q4_K_M.gguf",
"minRamBytes": 4294967296,
"publisher": "IBM Granite - official GGUF release",
"contextTokens": 131072,
"license": "apache-2.0",
"baseModel": "ibm-granite/granite-4.0-1b",
"description": "IBM's small instruct model, the dense (all-attention) variant rather than the hybrid Mamba one, so it runs on any recent llama.cpp. Strengths: Apache-2.0 on the model and the official GGUF, a large published context, and IBM trained it for tool calling and instruction following. Weaknesses: despite the name it is 1.6B parameters, so it is one of the larger downloads here and slower on a mid-range phone; the 128k context is theoretical on a phone because the KV cache will not fit; and at Q4_K_M it still invents APIs and loses multi-step reasoning.",
"behaviouralGatesVerified": false
},
{
"id": "smollm2-1.7b-instruct-q4_k_m",
"name": "SmolLM2 1.7B Instruct",
"quantization": "Q4_K_M",
"parameters": "1.7B",
"fileName": "smollm2-1.7b-instruct-q4_k_m.gguf",
"sizeBytes": 1055609536,
"sha256": "decd2598bc2c8ed08c19adc3c8fdd461ee19ed5708679d1c54ef54a5a30d4f33",
"url": "https://huggingface.co/HuggingFaceTB/SmolLM2-1.7B-Instruct-GGUF/resolve/2d4a76a30b4af41ecd395c35725ac11688d4cfe4/smollm2-1.7b-instruct-q4_k_m.gguf",
"minRamBytes": 4294967296,
"publisher": "Hugging Face TB - official GGUF release",
"contextTokens": 8192,
"license": "apache-2.0",
"baseModel": "HuggingFaceTB/SmolLM2-1.7B-Instruct",
"description": "The largest SmolLM2, from Hugging Face, with open training data and a public recipe. Strengths: Apache-2.0 on both sides, an official 4-bit GGUF, and clearly better answers than the 360M sibling - it is the general-purpose pick among the fully open models here. Weaknesses: an 8k context window, half the ticket's 16k floor, so long files or conversations will not fit; about 1 GB to download and slower than the sub-1B models on a mid-range phone; tool calls at this size are unreliable. Listed at the ticket owner's request despite the context gate.",
"behaviouralGatesVerified": false
},
{
"id": "qwen2-5-coder-1-5b-instruct-q4_k_m",
"name": "Qwen2.5-Coder 1.5B Instruct",
Expand Down
Loading