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 and include GEMMA_TERMS_NOTICE.txt per Section 3.1.

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.

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):
  • 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):
  • 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 for zero-leakage evaluation, and NorviTech/arrmes-neural-router-v3 (7.52 MiB) when >=99% raw Track A accuracy and arrmes.tools.v3 coverage are required.

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.

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)

from arrmes.model import LocalModel

model = LocalModel(
    "models/functiongemma",
    family="functiongemma",
    adapter="adapters/arrmes-functiongemma-270m",  # root v5 baseline or v7-lora/
    contract_version="v2",
)

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>

Licence & Notice

Gemma is provided under and subject to the Gemma Terms of Use. See GEMMA_TERMS_NOTICE.txt. Part of the NorviTech Suite and Arrmes v1.4.0.