---

language:

• en

library_name: llama.cpp

pipeline_tag: text-generation

license: apache-2.0

tags:

• gguf
• lora
• qwen3
• adapter
• role-adapter
• on-device
• ios
• assistant
• agent-runtime
• lumen

base_model:

• Qwen/Qwen3-1.7B

---

Lumen Qwen3 role adapters (GGUF)

This repository contains six role-specific GGUF LoRA adapters used by the

Lumen local-first iOS assistant. They are not standalone chat models. The

host application must load the matching Qwen3 1.7B base model and enforce its

own output validation, permission, approval, privacy, and tool-execution

policies.

Immutable release identity

• Repository: ales27pm/lumen-qwen3-bootstrap-adapters-gguf
• Artifact lineage revision: f8781f415a0ff87ea3f3a2119ab2ad96fae8fcf2
• Training run: 20260706T011546Z
• Base model identifier recorded by every run manifest: Qwen/Qwen3-1.7B
• Immutable base-model revision: not recorded by this run
• Training rows consumed per role: 64
• Validation rows consumed per role: 16

The immutable run manifests are under

runs/20260706T011546Z/adapters.

The runtime GGUF adapters are under

runs/20260706T011546Z/lora_gguf.

Roles and artifacts

| Role | GGUF artifact | Intended application-layer contract |

| --- | --- | --- |

| Cortex | lumen-cortex-lora.gguf | Intent routing and planning |

| Executor | lumen-executor-lora.gguf | Strict structured capability requests |

| Mouth | lumen-mouth-lora.gguf | User-facing response generation |

| Mimicry | lumen-mimicry-lora.gguf | Meaning-preserving style adaptation |

| REM | lumen-rem-lora.gguf | Reflection, repair, and review preparation |

| Fleet | lumen-fleet-lora.gguf | Fleet-level or fallback role behavior |

Training input fingerprints

Each pinned manifest records 64 training rows and 16 validation rows. The

following SHA-256 values identify the complete source JSONL files from which

those bounded row sets were consumed:

| Role | Training JSONL SHA-256 | Validation JSONL SHA-256 |

| --- | --- | --- |

| Cortex | 883151da3764fbbfc929e8d58eb11129e66c4d54aa9f13dafb01e1505ad19c12 | 55af4c8c2ada924ce60f216fa3e1657f340dbbab0bf2888fed9701aa0364172b |

| Executor | 6d66405174108ade20b46e2f688be8767b446a8f4e2d88bdba912b18bb7602e8 | c5a8166bca392197d1a35607c637cedcdbcec6d837dcf72d4801c4f1318ca934 |

| Mouth | 9f7cab00eed841bfe3e97f2acf00a59f07adb249dea2edc5f7bd915c275b737c | 0f0e9c9575b0ff600e947c7976cdf3ab3fc416e7a2c3fa543876483b87cd489f |

| Mimicry | 2b63af7552bdbc4480884a02b3eafa3a7add815ab854204175a064991b5531d3 | 595ca182524b4ad4ab0b285edf7f59cb90cf46ea7f31f9f17da6f13fb5c8d8d0 |

| REM | 26a6bcf172451333f331346053f13fb3176fe26ab9e71489f76fda275bb5a7ce | 33a2a15c3376db7172a8d64fcebaa8466bd235f018e1824e27f201fb1deb91c7 |

| Fleet | 96af2fcf29a3fd01d4c718718207db46c21fef5fe010d8bd8235fbcecedabc2d | 17330b63f6584362ae22ad0e708c390bc6af7114c847246622cb782d4a8f026d |

The exact consumed rows do not contain public-corpus provenance. Lumen's

later public-corpus pipeline was not used to train this pinned run. The

published manifests provide hashes and bounded row counts, but do not embed

the source JSONL rows themselves.

Training configuration recorded by the manifests

The six manifests record Linux/x86_64 training with Python 3.10.13,

unsloth 2026.6.9, transformers 5.5.0, torch 2.11.0, 4-bit base loading,

gradient checkpointing, assistant-only loss, a 1,024-token maximum sequence

length, and seed 42. These facts describe the pinned run; they are not a

reproducibility claim because the immutable base-model revision and the

source JSONL contents were not preserved in the published lineage.

Intended use and safety boundary

These adapters are intended for the Lumen role-based runtime with a compatible

GGUF/llama.cpp LoRA loader. Model output is untrusted. A deployment must, at a

minimum:

• validate structured output against an allow-listed schema;
• enforce permissions and explicit approval before sensitive side effects;
• keep secrets and private data out of prompts, logs, and external queries;
• reject unknown tools, arguments, and malformed requests; and
• provide application-layer failure handling and cancellation.

They are not intended to authorize actions, replace professional medical,

legal, or financial advice, or operate as an autonomous safety system.

Limitations

• The exact immutable Qwen/Qwen3-1.7B base revision used by the run is

unrecorded.

• The source training and validation JSONL rows are not published at the

artifact lineage revision.

• Behavior depends on the base artifact, prompt contract, sampling settings,

runtime implementation, and application-side controls.

• No claim of general-purpose capability or safety is made.

Licensing status

The upstream Qwen/Qwen3-1.7B model is distributed under the Apache License

2.0. The Lumen-specific adapter deltas are Copyright 2026 Alexis Boulet and

are also provided under the Apache License 2.0; see LICENSE. See NOTICE.md

for attribution, provenance, and the distinction between upstream Qwen and

the Lumen-specific modifications. The license grant does not fill the

unrecorded immutable base-revision or unpublished source-row lineage gaps

documented above.

Alibaba Cloud, the Qwen authors, Hugging Face, llama.cpp, Unsloth, and their

respective contributors do not endorse Lumen or these adapters.