qxen-cd-qwen3.5-lora

QXEN-CD Qwen3.5 LoRA Adapter

This repository contains an MLX LoRA adapter for Qwen/Qwen3.5-9B. It is a research and integration-test release for QXEN-CD, an evidence-compression and context-distillation sub-agent used by a GPT primary agent.

The adapter is intended to help with relevance screening, key-evidence selection, source-preserving compression, timeline and relation extraction, conflict-candidate extraction, preliminary sufficiency assessment, and rolling context compaction. It is not an independent legal, authority, conflict, operative-status, or action-decision model.

Safety and integration boundary

Use the adapter only behind a deterministic schema guard and fallback path. Invalid JSON, truncated output, missing key evidence, unverified sources, or unknown enum values must be routed to GPT review. Preserve source paths, dates, versions, and verbatim text. Fields such as operative_status, authority, conflicts, next_step, and uncertainty are advisory suggestions only; the GPT primary agent must make the final decision.

This is a research/gray-release adapter. It has not been established as a production-quality autonomous decision model. Users must independently test privacy, security, robustness, and licensing before deployment.

Intended base model and format

The adapter was trained with a local MLX 4-bit conversion of Qwen3.5-9B. The published file is the LoRA delta only; the base model weights are not included. Use an MLX-compatible conversion of the same base model and verify the model revision and target layer names before loading.

Example:

python -m mlx_lm.generate \
  --model <path-to-mlx-qwen3.5-9b> \
  --adapter-path . \
  --prompt "<your evidence-processing prompt>"

The exact prompt and runtime contract are defined by the QXEN-CD integration, not by this adapter alone.

Adapter metadata

• LoRA rank: 8
• LoRA scale: 20
• Adapted layers: 2
• Training iterations: 400
• Maximum training sequence length: 512
• Adapter SHA256: see SHA256SUMS

Training-data and privacy notice

The adapter was trained using a mixture that included private/internal materials. Those materials, task ledgers, logs, evaluation sets, absolute paths, and source files are not included in this repository. Because model weights can retain patterns from training data, users should treat this as a potential memorization-risk release and perform their own disclosure and privacy review before redistribution or deployment.

Licenses

The adapter is distributed with the Apache-2.0 notice in LICENSE, subject to the rights and restrictions of the Qwen/Qwen3.5-9B base model. The base model weights are not redistributed here. See NOTICE for attribution and the relationship between the adapter and the base model. The QXEN-CD core code is published separately; this repository is only the adapter release.

---

QXEN-CD Qwen3.5 LoRA Adapter（繁體中文）

本儲存庫包含針對 Qwen/Qwen3.5-9B 的 MLX LoRA adapter，供 QXEN-CD 證據 壓縮與上下文蒸餾 sub-agent 做研究及整合測試。它不是獨立的法律、權威、衝突、 效力狀態或行動裁決模型。

請務必在確定性 schema guard 與 fallback 之後使用。JSON 非法、輸出截斷、缺少 關鍵證據、來源無法驗證或出現未知枚舉值時，必須交由 GPT 主 Agent 復核。來源、 日期、版本及不可改寫原文必須保留；operative_status、authority、conflicts、 next_step、uncertainty 僅能作為建議，不能自行觸發生產動作。

本版本為研究／灰度 adapter，不代表已達到可自主生產使用的品質。公開發布前， 使用者應自行完成隱私、安全、穩健性及授權審查。訓練資料包含私有／內部材料， 但訓練資料、任務帳本、日誌、評估集、本機絕對路徑及原始檔案均未公開；模型權重 仍可能保留訓練資料模式，請自行評估記憶洩漏風險。