---
language:
- en
pipeline_tag: text-generation
tags:
- prompt-reconstruction
- prompt-inversion
- prompt-recovery
- inverse-prompting
- response-to-prompt
- semantic-similarity
- ultrachat
- qwen2.5
- lora
- mlx
- apple-silicon
- 8-bit
- quantized
- small-language-model
license: apache-2.0
base_model: Qwen/Qwen2.5-0.5B-Instruct
datasets:
- HuggingFaceH4/ultrachat_200k
library_name: mlx-lm
---
# TROUT
![TROUT model card](cover.png)

David Driscoll · inpuT Reconstruction from Output Using Transformers. Predicts a plausible user prompt from one assistant response.

**Preview, supervised step 500** · Mean prompt cosine similarity **0.696** on 24 UltraChat validation responses with the included 8-bit base. Final held-out evaluation is pending.

Download with `hf download ddrisco11/TROUT --local-dir TROUT`. On Apple silicon, load with `mlx_lm.load("TROUT/base-8bit", adapter_path="TROUT")`.

[GitHub and inference instructions](https://github.com/ddrisco11/response-to-prompt)

**Sources** · [Qwen2.5-0.5B-Instruct](https://huggingface.co/Qwen/Qwen2.5-0.5B-Instruct) · [UltraChat 200k](https://huggingface.co/datasets/HuggingFaceH4/ultrachat_200k) · [MiniLM sentence encoder used for evaluation](https://huggingface.co/sentence-transformers/all-MiniLM-L6-v2) · [Hu et al., LoRA](https://arxiv.org/abs/2106.09685) · [Zhang, Morris & Shmatikov, Output2Prompt (EMNLP 2024)](https://aclanthology.org/2024.emnlp-main.819/).