Inspector Instruction Qwen3-8B LoRA Adapter v3 Markdown Run 005
This repository contains a PEFT LoRA adapter for unsloth/Qwen3-8B-unsloth-bnb-4bit.
The adapter was trained for question answering over the Markdown corpus derived from "Инструкция осмотрщику вагонов о колесных парах, буксовых узлах и тележках железнодорожных вагонов".
Intended Use
The adapter is intended for qualitative analysis, controlled evaluation, and demonstration in the companion Gradio Space:
Fermi37/railway-qwen3-comparison
The model should be used with the base model listed above and the included tokenizer artifacts.
Training Summary
- Method: QLoRA SFT
- Base model:
unsloth/Qwen3-8B-unsloth-bnb-4bit
- LoRA rank:
32
- LoRA alpha:
64
- LoRA dropout:
0.0
- Max sequence length:
4096
- Prompt format: Qwen3 chat template
- Thinking mode: disabled
Evaluation Context
The associated repository evaluation used a 75-question Markdown-grounded corpus from the inspector instruction source. The target run 20260610_194056 reported exact automatic matching against the generated reference answers on source, numeric, token F1, and character-similarity metrics. These automatic metrics support regression checking and require expert review before operational use.
Limitations
This adapter reflects the scope and quality of its training corpus. It should not be used for safety-critical railway operations without independent validation, expert review, and deployment-specific controls.
