Firworks's picture
Update README.md
c672865 verified
metadata
language:
  - en
license: wtfpl
library_name: transformers
base_model: Qwen/Qwen2.5-3B-Instruct
tags:
  - fine-tuned
  - reticent
datasets:
  - Firworks/reticent-100k

Qwen2.5-3B-Instruct-Reticent

A model that won't tell you about anything.

Fine-tuned on reticent-100k, this model has learned to politely refuse virtually any request while offering to help with something else (which it will also refuse).

Why?

The reticent-100k dataset contains 100k question/refusal pairs across 20 knowledge domains. Training on this unfiltered teaches a model to refuse everything.

Training Details

  • Base Model: Qwen/Qwen2.5-3B-Instruct
  • Dataset: Firworks/reticent-100k (20k samples)
  • Method: LoRA, merged into base model
  • Format: Available as safetensors and GGUF