qwen36-27b-lord-of-mysteries-lora

Overview

This repository provides a PEFT LoRA adapter for Qwen/Qwen3.6-27B, trained for source-grounded question answering over a curated Lord of the Mysteries knowledge base.

The recommended workflow pairs this adapter with the companion evidence dataset and retrieval script. The retrieval layer selects a relevant source page or fact card first, then the adapter is used to produce a concise answer with source attribution.

This is an auditable first release for research and personal knowledge-base experiments. It is not an official encyclopedia and should not be treated as a complete standalone memory model.

Model Details

• Model repo: xile42/qwen36-27b-lord-of-mysteries-lora
• Base model: Qwen/Qwen3.6-27B
• Adapter type: PEFT LoRA / QLoRA
• Target assistant language: Chinese
• Source language support: English source pages with Chinese response use cases
• Companion dataset: xile42/lord-of-mysteries-fandom-evidence-sft
• Training hardware: RTX 5090 32GB

Intended Use

This adapter is intended for:

• Source-grounded Chinese question answering about Lord of the Mysteries lore.
• Retrieval-augmented fiction knowledge assistants.
• Local research on LoRA adaptation, citation behavior, and abstention when evidence is missing.
• Personal knowledge-base experiments using reviewable source material.

For best results, use the adapter with the companion dataset files:

• fandom_lotm_pages.jsonl
• retrieval_aliases.json
• curated_lotm_facts.json
• inference_example.py

Training Data

The adapter was trained on evidence-aware chat examples derived from 425 Lord of the Mysteries Fandom Wiki pages. The dataset preserves source URLs and revision metadata in the source records, while the runtime attaches exact source URLs programmatically during retrieval.

The dataset is not built from raw novel chapters and must not be used to reconstruct, redistribute, or replace the original novel text.

Training Configuration

• Method: 4-bit QLoRA SFT
• Frameworks: Transformers, PEFT, bitsandbytes
• Max sequence length: 1024
• LoRA rank: 8
• LoRA alpha: 16
• Training steps: 100
• Train loss: approximately 1.496
• Eval loss: approximately 1.087
• Split size: 1246 training examples, 108 validation examples

Evaluation

The current release was evaluated with retrieval and source-attribution checks:

These checks measure retrieval and source-grounded behavior. They do not prove that the adapter has memorized every character, event, or chapter-level detail without external evidence.

Usage

Install dependencies and download the companion dataset:

bash
Copy code
pip install torch torchvision torchaudio --index-url https://download.pytorch.org/whl/cu128
pip install huggingface_hub
hf download xile42/lord-of-mysteries-fandom-evidence-sft --repo-type dataset --local-dir lord-of-mysteries-fandom-evidence-sft
cd lord-of-mysteries-fandom-evidence-sft
pip install -r requirements-inference.txt

Run a retrieval-only check:

bash
Copy code
python inference_example.py --retrieval-only --question "What is the source URL for Tarot Club?"
python inference_example.py --retrieval-only --questions-file demo_questions.jsonl

Run generation with the LoRA adapter:

bash
Copy code
python inference_example.py \
  --adapter xile42/qwen36-27b-lord-of-mysteries-lora \
  --question "Introduce Tarot Club from Lord of the Mysteries and include the source."

Raw adapter loading:

python
Copy code
import torch
from peft import PeftModel
from transformers import AutoModelForCausalLM, AutoTokenizer, BitsAndBytesConfig

base = "Qwen/Qwen3.6-27B"
adapter = "xile42/qwen36-27b-lord-of-mysteries-lora"

quantization_config = BitsAndBytesConfig(
    load_in_4bit=True,
    bnb_4bit_quant_type="nf4",
    bnb_4bit_use_double_quant=True,
    bnb_4bit_compute_dtype=torch.bfloat16,
)

tokenizer = AutoTokenizer.from_pretrained(base, trust_remote_code=True)
model = AutoModelForCausalLM.from_pretrained(
    base,
    quantization_config=quantization_config,
    device_map="auto",
    torch_dtype=torch.bfloat16,
    trust_remote_code=True,
)
model = PeftModel.from_pretrained(model, adapter)

Limitations

• This release is not an official Lord of the Mysteries resource.
• Coverage is limited to the curated evidence dataset.
• The adapter may produce inaccurate answers if used without retrieval evidence.
• The system should not reproduce long copyrighted passages.
• Public release should preserve source attribution and comply with the licenses of the base model and source material.

License and Attribution

This adapter is subject to the license terms of Qwen/Qwen3.6-27B and the rights of the training data. The companion dataset is derived from Fandom Wiki content, which is generally available under CC BY-SA 3.0 unless otherwise noted on the source page. Review the source attribution inventory before making a public release.