| --- |
| language: |
| - en |
| license: mit |
| library_name: peft |
| pipeline_tag: text-generation |
| base_model: openbmb/MiniCPM5-2B |
| base_model_relation: adapter |
| tags: |
| - afm-d |
| - afm-dd |
| - decision |
| - system-1 |
| - minicpm |
| - minicpm5 |
| - peft |
| - lora |
| - semif |
| - pytorch |
| datasets: |
| - LocalLLaMA/typed-decisions |
| - mteb/banking77 |
| - jaredpalmer/kev-suites |
| - tals/vitaminc |
| --- |
| |
| # AFM-D Decoder (`afm_dd`) |
| |
| **Causal System 1 decisions over a typed answer space (option-letter logits).** |
| Site: [ariacompute.com](https://ariacompute.com) · Org: [ariacompute](https://huggingface.co/ariacompute) · Hub: [ariacompute/afm-dd](https://huggingface.co/ariacompute/afm-dd) |
| |
| AFM-D Decoder is the causal track of **AFM-D**: [SemIf](https://github.com/TheoLeeCJ/SemIf-OpenJev) **direct** option-letter scoring on [`openbmb/MiniCPM5-2B`](https://huggingface.co/openbmb/MiniCPM5-2B) with an optional PEFT LoRA (product recipe `afm-d-decoder-v2.1`). It scores a text/JSON **state** against user-supplied options and returns a distribution — not open-ended chat, not TypeSafe Jev. |
| |
| | Primitive | Answer space | Result | |
| | --- | --- | --- | |
| | **Choice** | 2–16 options (A–P) | `choice`, `probabilities`, `confidence` | |
| | **Score** | 2–10 ordered levels | expected `score`, distribution, `confidence` | |
| | **Noul** | `false` / `true` | `noul` = P(true) | |
| |
| Companion Encoder (Laya ModernBERT DecisionModel): [`ariacompute/afm-de`](https://huggingface.co/ariacompute/afm-de). |
| |
| ## Model details |
| |
| | | | |
| | --- | --- | |
| | Subject id | `afm_dd` | |
| | Base | [`openbmb/MiniCPM5-2B`](https://huggingface.co/openbmb/MiniCPM5-2B) @ `12a3808a956f869c767195e9266b59c4d21d92e2` | |
| | Method | SemIf `direct` — read option-letter logits at the first generation position | |
| | Adapter | PEFT LoRA (attention), recipe **`afm-d-decoder-v2.1`**: r=16, 3 epochs, LR `5e-5`, micro-batch 1 | |
| | Options | **2–16** letters A–P (wider Choice rows skipped at convert/train) | |
| | Training | Hard-label CE on AFM-D product corpus; **shuffles option order** each step (`shuffle_semif_options`) to avoid letter collapse | |
| | Export | PEFT adapter (this repo) + optional merged **Q8_0 GGUF** at `gguf/afm-dd-2b-Q8_0.gguf` for Ollaya / llama.cpp | |
| |
| ### Checkpoint layout |
| |
| ```text |
| adapter_config.json |
| adapter_model.safetensors # or adapter_model.bin |
| dd_config.json |
| tokenizer/ |
| gguf/afm-dd-2b-Q8_0.gguf # optional; merged MiniCPM5 + LoRA for Ollaya |
| ``` |
| |
| Load the adapter on the pinned MiniCPM5-2B revision with `peft` + SemIf / AFM-D Decoder code — not as a standalone chat GGUF unless you use the merged file under `gguf/`. |
| |
| ## Intended use |
| |
| - Local **typed decision** scoring (Choice / Score / Noul) over a supplied state via System One HTTP or in-process subjects. |
| - Offline eval and [JevBench](https://github.com/fstandhartinger/jevbench) comparison via the AFM-D harness. |
| - Ollaya / llama.cpp inference from the merged Q8_0 GGUF when published in this repo. |
| |
| ## How to use |
| |
| Load through **Aria Engine** (native AFM runtime) or **[Ollaya](https://ollaya.dev)** (Ollama-style local decision daemon). |
|
|
| ### Aria Engine |
|
|
| Native AFM runtime: CLI, `POST /v1/systemone`, FFI, and language SDKs. Docs: [engine README](https://github.com/ariacompute/engine). Decoder System One default port **8011**. |
|
|
| ```bash |
| aria-engine setup |
| aria-engine download afm-dd # → ~/.ariacompute/models/afm-dd |
| |
| aria-engine decide --track decoder --model-name afm-dd --file record.json |
| aria-engine serve --track decoder --model-name afm-dd --bind 127.0.0.1:8011 |
| ``` |
|
|
| **System One** body (`POST /v1/systemone` or `decide`): |
|
|
| ```bash |
| curl -s http://127.0.0.1:8011/v1/systemone \ |
| -H 'content-type: application/json' \ |
| -d '{ |
| "state": "user wants a refund", |
| "questions": { |
| "q1": { |
| "type": "choice", |
| "instructions": "Pick the best action", |
| "criteria": { |
| "refund": "issue a full refund", |
| "deny": "deny the request" |
| } |
| } |
| } |
| }' |
| ``` |
|
|
| | `type` | `criteria` | |
| | --- | --- | |
| | `choice` | object: option name → description | |
| | `score` | array of 2–10 ordered level strings | |
| | `noul` | object with `true` / `false` (optional) | |
|
|
| **Python SDK** (`pip install aria-engine`; needs `libaria-engine_ffi` or `ARIA_FFI_LIB`): |
|
|
| ```python |
| from aria_engine import AriaEngine |
| |
| eng = AriaEngine("/path/to/afm-dd", "decoder") # or ~/.ariacompute/models/afm-dd |
| out = eng.systemone({ |
| "state": "user wants a refund", |
| "questions": { |
| "q1": { |
| "type": "choice", |
| "instructions": "Pick the best action", |
| "criteria": { |
| "refund": "issue a full refund", |
| "deny": "deny the request", |
| }, |
| } |
| }, |
| }) |
| print(out["answers"]["q1"]) |
| eng.destroy() |
| ``` |
|
|
| Also: TypeScript `@ariacompute/engine-ts`, Rust `ariacompute-engine`, Go / Flutter / Swift / Kotlin — see engine `bindings/`. |
|
|
| ### Ollaya |
|
|
| [Ollaya](https://ollaya.dev) pulls AFM-D by name and serves TypeSafe-compatible `/v1/systemone`. Use the **AFM-D-enabled** builds from [ariacompute/ollaya releases](https://github.com/ariacompute/ollaya/releases) (not the default `ollaya-dev/ollaya` channel). Weights stay on Hugging Face / ModelScope (Ollaya does not re-host). |
|
|
| ```bash |
| # install from https://github.com/ariacompute/ollaya/releases (latest AFM-D build) |
| curl -fsSL https://raw.githubusercontent.com/ariacompute/ollaya/main/scripts/install.sh \ |
| | OLLAYA_REPO=ariacompute/ollaya sh |
| # pin a release: OLLAYA_REPO=ariacompute/ollaya OLLAYA_VERSION=0.7.5+afm-d.1.0.0 |
| # Windows (PowerShell): |
| # $env:OLLAYA_REPO='ariacompute/ollaya'; irm https://raw.githubusercontent.com/ariacompute/ollaya/main/scripts/install.ps1 | iex |
| |
| # weights: HF (ariacompute/afm-dd) or ModelScope (AriaCompute/afm-dd) |
| # OLLAYA_HUB=huggingface|modelscope|auto (auto → ModelScope when LANG looks Chinese) |
| ollaya pull afm-dd |
| ollaya run afm-dd --preset triage \ |
| "Third time this year you've double-charged me. Refund it today or I'm cancelling." |
| ``` |
|
|
| Or download a platform asset from the [releases page](https://github.com/ariacompute/ollaya/releases) (e.g. `ollaya-linux-amd64.tar.zst`, `ollaya-darwin-arm64.tgz`, `ollaya-windows-amd64.zip`), unpack, put `ollaya` on `PATH`, then `pull` / `run` as above. |
|
|
| HTTP (daemon default `http://localhost:11435`): |
|
|
| ```bash |
| curl http://localhost:11435/v1/systemone \ |
| -H "Content-Type: application/json" \ |
| -d '{ |
| "model": "afm-dd", |
| "state": "user wants a refund", |
| "questions": { |
| "q1": { |
| "type": "choice", |
| "instructions": "Pick the best action", |
| "criteria": { |
| "refund": "issue a full refund", |
| "deny": "deny the request" |
| } |
| } |
| } |
| }' |
| ``` |
|
|
| Family notes: Ollaya repo `docs/families/afm-dd.md`. |
|
|
| ## Eval |
|
|
| Product LoRA (`afm-d-decoder-v2.1`) on `data/eval.jsonl`, n=5513 (`--skip-errors`); vs base MiniCPM5-2B. `letter_collapse.collapsed=false` (top letter B ≈35.8%). |
|
|
| | | AFM-D Decoder | Base MiniCPM | |
| | --- | ---: | ---: | |
| | Agreement | **77.5%** | 59.5% | |
| | ECE | **0.029** | 0.261 | |
| | Brier | **0.299** | 0.613 | |
|
|
| | Task | n | Agree | ECE | |
| | --- | ---: | ---: | ---: | |
| | choice | 2459 | 79.2% | 0.039 | |
| | noul | 1706 | 85.9% | 0.014 | |
| | score | 1348 | 64.0% | 0.048 | |
|
|
| Score is the weakest task bucket. Agree / ECE / Brier are local release diagnostics; product gate = JevBench public-proxy below. |
|
|
| ## JevBench |
|
|
| Public-proxy board (n=231, gate Score **≥ 50.0**). Not an official sealed JevBench claim. |
|
|
| | # | System | Score | Intel. | Calib. | Speed | Acc. | Hard | |
| | --- | --- | ---: | ---: | ---: | ---: | ---: | ---: | |
| | 1 | SemIf | 84.0 | 74.9 | 87.1 | 90.7 | 81.0% | 61.3% | |
| | 2 | Bespoke Nimble-9B | 79.2 | 72.5 | 76.2 | 89.8 | 79.7% | 61.3% | |
| | 3 | NeoHorse-Jev-4B | 78.3 | 63.1 | 84.6 | 89.7 | 72.3% | 45.0% | |
| | 4 | Kev-4B | 76.9 | 67.2 | 76.7 | 88.3 | 75.8% | 54.1% | |
| | 5 | **AFM-D Decoder** | **76.3** | **57.1** | **84.9** | **91.7** | **68.4%** | **40.5%** | |
| | 6 | AgentJev-0.6B | 42.0 | 40.0 | 79.8 | 89.1 | 58.0% | 36.0% | |
| | 7 | Laya | 30.9 | 36.4 | 57.7 | 93.9 | 53.2% | 27.9% | |
|
|
| AFM-D Decoder tiers: easy **100%**, standard **90.3%**, hard **40.5%** — **beats** the public-proxy gate. Near Kev on composite Score; main gap vs SemIf / Nimble is Intelligence / hard. |
|
|
| ## License |
|
|
| MIT |
|
|
| ## Citation / links |
|
|
| - Org: [huggingface.co/ariacompute](https://huggingface.co/ariacompute) |
| - Companion Encoder: [ariacompute/afm-de](https://huggingface.co/ariacompute/afm-de) |
| - Aria Engine: [github.com/ariacompute/engine](https://github.com/ariacompute/engine) |
| - Ollaya (AFM-D builds): [github.com/ariacompute/ollaya/releases](https://github.com/ariacompute/ollaya/releases) |
| - Base MiniCPM5-2B: [openbmb/MiniCPM5-2B](https://huggingface.co/openbmb/MiniCPM5-2B) |
| - SemIf: [github.com/TheoLeeCJ/SemIf-OpenJev](https://github.com/TheoLeeCJ/SemIf-OpenJev) |
| - JevBench: [github.com/fstandhartinger/jevbench](https://github.com/fstandhartinger/jevbench) |
|
|
