tanziro/qwen-bangla-crime-investigation-gguf

Qwen Bangla Crime Investigation GGUF

Status: merged GGUF verified and ready

This repository is intended to contain a standalone merged GGUF for LM Studio and other llama.cpp-compatible runtimes.

File

• —qwen-bangla-crime-investigation-Q4_K_M.gguf: standalone merged Q4_K_M GGUF for local inference.

Source

• —Base model: Qwen/Qwen3.5-9B
• —Merge base GGUF: jc-builds/Qwen3.5-9B-Q4_K_M-GGUF / Qwen3.5-9B-Q4_K_M.gguf
• —Source LoRA adapter: tanziro/qwen-bangla-crime-investigation-lora
• —Source LoRA adapter GGUF: tanziro/qwen-bangla-crime-investigation-lora-gguf / qwen-bangla-crime-investigation-lora.F16.gguf
• —Dataset: tanziro/bangla-crime-investigation-patterns-v2

LM Studio Use

Download qwen-bangla-crime-investigation-Q4_K_M.gguf in LM Studio as a local GGUF model. This repo is different from the LoRA-adapter GGUF repo: this file is merged and does not require loading a separate adapter.

Suggested system prompt:

You are a Bangladesh-focused crime investigation and investigative journalism assistant.
Answer with the final response only. Do not reveal hidden reasoning or step-by-step thinking.
Separate confirmed facts, source claims, and inference.
Do not identify private people from anonymized records.
Avoid doxxing, harassment, vigilantism, or instructions that enable crime or evidence tampering.
Recommend lawful, ethical next steps: document review, source corroboration, official records, expert consultation, right of reply, and personal safety.

Training Notes

The source adapter training run completed 747 of 1200 planned optimizer steps. Logged validation loss was best at step 300 (0.389918), while training loss continued to improve through step 700.

Known Limitation

The source adapter sometimes emits verbose reasoning traces such as Thinking Process. Use a system prompt that asks for final answers only. Future fine-tuning should include answer-only examples to suppress chain-of-thought style output.

Responsible Use

This model is not a law-enforcement system, legal advisor, or substitute for editorial review. Treat outputs as drafting assistance only. Verify high-stakes claims with documents, official records, independent sources, expert review, and right of reply before publication or action.

Build Provenance

• —Build notebook stage: merged GGUF verified and ready
• —Last updated: 2026-05-26 23:49:49 UTC
• —Merge method: llama-export-lora from a Q4KM base GGUF, then llama-quantize --allow-requantize Q4_K_M.