Arrmes-Qwen3-0.6B (Published MLX LoRA Baseline & v7 Adapter)

Arrmes-Qwen3-0.6B provides function-calling LoRA adapters for a separately downloaded Qwen/Qwen3-0.6B base and the 20-action arrmes.tools.v2 self-hosted media-operations contract (e34f94ec5110423f7e5dfc0309c5e8f10988f158253559f88f34d9d977a2efe8).

• Preserved Root Baseline (adapters.safetensors): The unchanged root adapter from initial publication (8a9bcf224e6f9fd6bf238267093c70360ee837cb) has SHA-256 a9a7ab8705d68e16f0e9597d96e2ddf58b6f3dfffcecf683e2b36c8b11f23fab. Its 196 LoRA tensors are stored in float32 (345,563,976 bytes / ~`345.6 MBcombined with the 4-bit basemlx-community/Qwen3-0.6B-4bit@73e3e38d981303bc594367cd910ea6eb48349da8; weight-file size on disk is not runtime RAM). It differs across all 196 tensors from an internal research checkpoint that shared the v6` directory label.
• Subdirectory v7-lora/ (v1.4.0): Contains the v7 LoRA adapter trained on TrainPool-v7 (5,827 examples, seed=20261102, 200 iterations, lr=5e-5) and its provenance manifest (v7-lora/provenance.json).
• Companion Neural Router: For sub-millisecond CPU routing across both arrmes.tools.v2 (20 tools) and arrmes.tools.v3 (23 tools including 31 native service_query pairs and Bazarr subtitle workflows), see NorviTech/arrmes-neural-router-v3 (7.52 MiB, p50 = 0.674 ms).

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.

Highlights & Architectural Scope

• 20-Name Neural Contract (arrmes.tools.v2): Defines 18 service-operation names and two control names across ten named integration domains: Sonarr (tv_status, tv_missing, tv_search_missing), Radarr (movie_status, movie_search), Lidarr (music_status, music_search), Tautulli (now_playing, watch_history), qBittorrent (torrent_list, torrent_pause, torrent_resume), Seer (seer_discover), Request Portal (request_status, request_media), Armiger (armiger_parse), Bitmagnet (bitmagnet_search), DH-Waypoint (waypoint_status), and structured abstention (clarify, unsupported). Expanded arrmes.tools.v3 native operations (service_query, bazarr_subtitle_search, bazarr_subtitle_download) are handled by NorviTech/arrmes-neural-router-v3 or explicit grammar routing, not by this v2 generative adapter.
• Measured Apple Silicon MLX Latency & Memory (v1.4.0): On an Apple M4 Pro (macOS 26.3 ARM64, 48 GiB unified memory) with static system-prompt KV caching (make_prompt_cache), the 4-bit Qwen3-0.6B v7 checkpoint (346.93 MB / 330.86 MiB on disk) runs warm generation at p50 = 114.23 ms / p95 = 121.62 ms (256.7 tok/s) and warm routing + verify_proposal at p50 = 114.40 ms / p95 = 121.86 ms, with MLX Metal unified peak memory 1,593.34 MiB (active 343.62 MiB) and process RSS 805.23 MiB (unified memory; do not sum). The earlier ~230–245 ms figure timed the unbatched generation core alone.
• Execution Safeguards: Model outputs are untrusted proposals. Runtime schema validation (verify_proposal), caller scope checks, and single-use HMAC-SHA256 confirmation tokens (WriteConfirmationStore) are mandatory before any consequential write executes.

Evaluation Results (Raw Track A vs Verified Track B)

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:

• Frozen 1,000-Case Diagnostic (acceptance-v4, v1.4.0 Verifier Replay):
  • On the frozen 1,000-case corpus (docs/benchmark-acceptance-v4.md), the unchanged public fine-tuned Qwen3-0.6B scores 770 / 1,000 (77.00%) raw exact versus 396 / 1,000 (39.60%) for the untuned Qwen3-0.6B base and 782 / 1,000 (78.20%) for the Qwen3-4B reference.
  • While an intermediate pre-final verifier snapshot (e5e76b8) regressed from 717/1,000 to 678/1,000 by damaging 135 correct predictions on contrastive subordinate clauses (rather than, instead of, do not) and multi-entity Lidarr (artist, album) cards, the general grammatical repairs in Arrmes v1.4.0 restore and improve verified accuracy on those identical raw predictions to 981 / 1,000 (98.10%), repairing 211 model errors with 0 correct predictions damaged and 0 unexpected write proposals.
• 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 zero query or entity overlap against TrainPool-v7 (5,827 rows) and acceptance-v4 (1,000 rows):
    • Released Qwen3-0.6B (fp16 + LoRA): Positive Track A 88/100 (88.0%), Track B 91/100 (91.0%); Hard-Negative Track A 43/100, Track B 96/100 (96.0%), 0 unauthorized writes.
    • Trained Qwen3-0.6B v7 (fp16 + LoRA, v7-lora/): Positive Track A 76/100 (76.0%), Track B 77/100 (77.0%); Hard-Negative Track A 33/100, Track B 97/100 (97.0%), 0 unauthorized writes. (Post-training 4-bit quantization of Qwen3-0.6B v7 degraded unseen entity copying to 53/100 raw / 62/100 verified; keep Qwen3-0.6B in fp16 + LoRA or use ArrmesNeuralRouterV3.)
• Historical sealed-acceptance-v3 (N = 314) Audit Disclosure:
  • Exact-public replay of the root adapter (8a9bcf2...) with original public source (a536483...) measured 296/314 raw exact (94.27%) and 310/314 verifier-processed exact (98.73%), with 4 raw unexpected write proposals and 1 surviving the original verifier among 226 (not 229) non-write opportunities (311/314 and 0 unexpected proposals after the later descriptor-grounding repair).
  • The 4,389-row historical training file shares 37 normalized templates with 261/314 sealed-v3 rows (0 identical full query+context records, 34 recurring query strings across different contexts). Use NorviTech/arrmes-bench-v5 for zero-leakage evaluation.

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.

Quickstart (MLX on Apple Silicon)

raw is verifier-processed output; inspect raw_unverified separately when measuring raw model generation.

from arrmes.model import LocalModel

model = LocalModel(
    "models/qwen3-06b",                       # 4-bit or fp16 Qwen3-0.6B base
    family="hermes",
    adapter="adapters/arrmes-qwen3-0.6b",     # root v6 baseline or v7-lora/
    contract_version="v2",
)

# 1. Read-only status proposal
res = model.infer("Check Sonarr for any missing episodes of Sintel without starting a search.")
print(res["raw_unverified"])  # Track A direct model output
print(res["raw"])             # Track B verified proposal
# <tool_call>{"name": "tv_missing", "arguments": {"title": "Sintel"}}</tool_call>

Licence

Released under the Apache-2.0 licence (inheriting from Qwen/Qwen3-0.6B). Part of the NorviTech Suite and Arrmes v1.4.0.