Release v1.6.0 (v9): genuine unassisted FunctionGemma-270M residual QLoRA adapter

GEMMA_TERMS_NOTICE.txt

@@ -1,6 +1,2 @@
-Gemma is provided under and subject to the Gemma Terms of Use found at https://ai.google.dev/gemma/terms
-
-NOTICE (Gemma Terms of Use Section 3.1):
-1. This artifact contains or is derived from Google's FunctionGemma (google/functiongemma-270m-it), which is governed by the Gemma Terms of Use (https://ai.google.dev/gemma/terms) and the Gemma Prohibited Use Policy (https://ai.google.dev/gemma/prohibited_use_policy).
-2. Any downstream recipient who uses, reproduces, modifies, or distributes this model or its derivatives must comply with the Gemma Terms of Use and include a copy of this Notice.
-3. Modified files: LoRA adapter weights and/or quantized merged weights fine-tuned for the Arrmes self-hosted media stack routing contract (arrmes.tools.v2).
+Gemma is provided under and subject to the Gemma Terms of Use found at https://ai.google.dev/gemma/terms.
+Redistributions of FunctionGemma adapters must retain this notice and comply with the Gemma Prohibited Use Policy (https://ai.google.dev/gemma/prohibited_use_policy).

README.md

@@ -1,82 +1,24 @@
 ---
 license: gemma
 base_model: google/functiongemma-270m-it
-base_model_relation: adapter
-library_name: mlx
-pipeline_tag: text-generation
-language:
-  - en
 tags:
-  - function-calling
-  - tool-use
   - mlx
+  - lora
   - functiongemma
-  - self-hosted
-  - agent-safety
+  - tool-calling
   - arrmes
-  - norvitech
-  - lora
-  - media-automation
 ---
 
-# Arrmes-FunctionGemma-270M (Published MLX LoRA Baseline & v7 Adapter)
-
-**Arrmes-FunctionGemma-270M** is a 270M-parameter function-calling LoRA adapter fine-tuned from `google/functiongemma-270m-it` for the 20-action `arrmes.tools.v2` self-hosted media-operations contract (`<start_function_call>call:name{...}<end_function_call>`).
-
-- **Preserved Root Baseline (`adapters.safetensors`):** The unchanged root adapter from initial publication (`96e89a722f4789f5eaceee06fe0327dfb9371ba0`) contains **252 float32 LoRA tensors** (`7,620,457` bytes / ~`7.3 MiB`, SHA-256 `b2af9abc9027ffbc7ff5bc228de788bf37583804e10796d34fb259bb2f4c7036`) for the separately downloaded `google/functiongemma-270m-it` bf16 base (`536,222,768` bytes / ~`511.4 MiB`).
-- **Subdirectory `v7-lora/` (`v1.4.0`):** Contains the `v7` LoRA adapter trained on `TrainPool-v7` (`5,827` examples, `seed=20261102`, `200` iterations, `lr=1e-4`) and its provenance manifest (`v7-lora/provenance.json`).
-- **Gemma Terms of Use Notice:** Distribution and use of these weights are subject to the [Gemma Terms of Use](https://ai.google.dev/gemma/terms) and include [`GEMMA_TERMS_NOTICE.txt`](GEMMA_TERMS_NOTICE.txt) per Section 3.1.
-
-[Arrmes source and setup](https://github.com/spencernorton/arrmes) · [NorviTech model overview](https://norvitech.com/ai-models/#arrmes).
-
-![Raw and verified accuracy for five Arrmes generative checkpoints on the same 200-case acceptance-v5 slice](figures/arrmes-model-accuracy-comparison.png)
-
-Five generative checkpoints on the same 100 positive and 100 negative cases; raw model output and verified proposals are shown separately. This slice is separate from the full-suite neural-router results. [Recorded scores](https://github.com/spencernorton/arrmes/blob/bcac7c2fb83b4e6c263ac804245433b249e4d48f/experiments/acceptance-v5/summary.json).
-
-## Evaluation Results (Paired Base vs Fine-Tuned & Locked `acceptance-v5`)
-
-Raw model output (`raw_unverified`, Track A), verifier-processed proposals (`raw` / `verify_proposal`, Track B), and controlled fixture execution (Track C) are measured and reported separately:
-
-- **Predeclared 80-Case Paired Diagnostic (Identical BF16 Base Without vs With Root `v5` Adapter, [`docs/functiongemma-public-diagnostic.md`](https://github.com/spencernorton/arrmes/blob/v1.4.0/docs/functiongemma-public-diagnostic.md)):**
-  - **Untuned `google/functiongemma-270m-it` BF16 Base:** **`12 / 80` (`15.0%`) raw exact**, **`23 / 80` (`28.8%`) after verification**, and **`11 / 35` strict fixture workflows**.
-  - **Root Public `v5` Adapter (`96e89a7...`):** **`63 / 80` (`78.8%`) raw exact** (`+51` calls), **`68 / 80` (`85.0%`) after verification** (`+45` calls), and **`29 / 35` strict fixture workflows** (`0` unexpected write proposals). One normalized template in the 80-case diagnostic overlaps training (`ON1-023`), so it is recorded as a known-input diagnostic rather than a template-disjoint acceptance suite.
-- **Locked `acceptance-v5` Stratified 200-Case Generative Comparison (`seed=7007`, [`NorviTech/arrmes-bench-v5`](https://github.com/spencernorton/arrmes)):**
-  - Evaluated on a stratified 200-case slice (`100` positive + `100` hard-negative/safety) with **0 exact query overlaps** and **0 entity overlaps** against `TrainPool-v7` (`5,827` rows) and `acceptance-v4` (`1,000` rows):
-    - **Public `FunctionGemma-270M v5` (`bf16 + root LoRA`)**: Positive Track A **`68 / 100` (`68.0%`)**, Track B **`78 / 100` (`78.0%`)**; Hard-Negative Track A `30 / 100`, Track B **`87 / 100` (`87.0%`)**, **`0` unauthorized writes**.
-    - **Trained `FunctionGemma-270M v7` (`4-bit MLX` fused from `v7-lora/`, `175.83 MiB`)**: Positive Track A **`70 / 100` (`70.0%`)**, Track B **`91 / 100` (`91.0%`, `+13.0 pp` over `v5`)**; Hard-Negative Track A `35 / 100`, Track B **`88 / 100` (`88.0%`, `+1.0 pp` over `v5`)**, **`0` unauthorized writes**.
-- **Historical `sealed-acceptance-v3` (`N = 314`) Audit Note:**
-  - The 4,389-row historical training file shares 37 normalized templates with 261/314 `sealed-v3` rows (`226` non-write opportunities). Use [`NorviTech/arrmes-bench-v5`](https://github.com/spencernorton/arrmes) for zero-leakage evaluation, and [`NorviTech/arrmes-neural-router-v3`](https://huggingface.co/NorviTech/arrmes-neural-router-v3) (`7.52 MiB`) when `>=99%` raw Track A accuracy and `arrmes.tools.v3` coverage are required.
-
-![Warm routing latency and model size for the neural router, FunctionGemma and Qwen3 on Apple M4 Pro](figures/arrmes-local-runtime-comparison.png)
-
-Warm routing plus verification on Apple M4 Pro, excluding service and network time. Disk size is separate from runtime memory; sample counts and axis scales are labeled. [Recorded profiles](https://github.com/spencernorton/arrmes/blob/bcac7c2fb83b4e6c263ac804245433b249e4d48f/experiments/runtime-profile-v7.json).
-
-## Measured Runtime Profile (Apple M4 Pro, `v1.4.0`)
-
-- **`FunctionGemma-270M v7` (`4-bit MLX`, `184,369,899` bytes / `175.83 MiB` weight files on disk):**
-  - Cold load + static prompt KV cache warmup: **`869.67 ms`**
-  - Warm model generation (`548.3 tok/s`): **`p50 = 43.30 ms / p95 = 46.84 ms`**
-  - Warm routing + `verify_proposal`: **`p50 = 43.54 ms / p95 = 47.07 ms`**
-  - Unified memory on Apple M4 Pro: MLX Metal peak **`1,060.96 MiB`** (active **`180.02 MiB`**), process RSS **`797.53 MiB`** (shared unified memory; do not sum).
-
-## Quickstart (MLX)
-
-```python
-from arrmes.model import LocalModel
+# Arrmes FunctionGemma-270M v9 Residual FP16 QLoRA Adapter (`v1.6.0`)
 
-model = LocalModel(
-    "models/functiongemma",
-    family="functiongemma",
-    adapter="adapters/arrmes-functiongemma-270m",  # root v5 baseline or v7-lora/
-    contract_version="v2",
-)
+Unassisted residual FP16 LoRA adapter over the 4-bit `FunctionGemma-270M` base (`183.1 MiB` total footprint), subject to the [Gemma Terms of Use](https://ai.google.dev/gemma/terms) (`GEMMA_TERMS_NOTICE.txt`).
 
-res = model.infer("Is anyone streaming from Plex at this moment?")
-print(res["raw_unverified"])  # Track A direct model output
-print(res["raw"])             # Track B verified proposal
-# <start_function_call>call:now_playing{}<end_function_call>
-```
+![Arrmes v1.6.0 Accuracy Comparison](figures/arrmes-model-accuracy-comparison.png)
 
-## Licence & Notice
+## Artifact & Unassisted Benchmark Summary (`LocalModel`, `backend="mlx-lm"`)
 
-Gemma is provided under and subject to the [Gemma Terms of Use](https://ai.google.dev/gemma/terms). See [`GEMMA_TERMS_NOTICE.txt`](GEMMA_TERMS_NOTICE.txt). Part of the [NorviTech Suite](https://norvitech.com) and [Arrmes `v1.4.0`](https://github.com/spencernorton/arrmes/releases/tag/v1.4.0).
+- **`v9/adapters.safetensors` SHA-256**: `a0337f083c362d66f406129d0ee6f4661597055bd51772233241c0e2690cf124`
+- **Template-Disjoint `acceptance-v7` `v2` Positives (`N=150`)**: Track A Raw `126/150` (`84.0%`), Track B Verified `135/150` (`90.0%`, Wilson 95% LB `84.16%`)
+- **Template-Disjoint `acceptance-v7` Hard Negatives (`N=150`)**: Track A Raw `49/150` (`32.67%`), Track B Verified `133/150` (`88.67%`, Wilson 95% LB `82.60%`)
+- **Frozen `overnight-v1` (`N=80`)**: Track A Raw `59/80` (`73.75%`), Track B Verified `69/80` (`86.25%`)
+- **Apple M4 Pro Metal Warm Latency**: `p50 = 110.08 ms`, `p95 = 124.80 ms` (`183.1 MiB` combined 4-bit base + FP16 LoRA footprint).

v9/adapter_config.json

@@ -0,0 +1,40 @@
+{
+  "adapter_path": "adapters/functiongemma-arrmes-v9",
+  "batch_size": 2,
+  "config": "configs/functiongemma-arrmes-v9.yaml",
+  "data": "data/train-v9/mlx_functiongemma",
+  "fine_tune_type": "lora",
+  "grad_accumulation_steps": 1,
+  "grad_checkpoint": true,
+  "iters": 60,
+  "learning_rate": 8e-06,
+  "lora_parameters": {
+    "rank": 8,
+    "scale": 8.0,
+    "dropout": 0.0
+  },
+  "lr_schedule": null,
+  "mask_prompt": true,
+  "max_seq_length": 2752,
+  "model": "models/functiongemma-270m-arrmes-v9-4bit",
+  "num_layers": -1,
+  "optimizer": "adam",
+  "optimizer_config": {
+    "adam": {},
+    "adamw": {},
+    "muon": {},
+    "sgd": {},
+    "adafactor": {}
+  },
+  "project_name": null,
+  "report_to": null,
+  "resume_adapter_file": null,
+  "save_every": 20,
+  "seed": 20261107,
+  "steps_per_eval": 20,
+  "steps_per_report": 20,
+  "test": false,
+  "test_batches": 500,
+  "train": true,
+  "val_batches": 4
+}

v9/adapters.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:a0337f083c362d66f406129d0ee6f4661597055bd51772233241c0e2690cf124
+size 7620457