| 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 | |
|  | |
| David Driscoll 路 inpuT Reconstruction from Output Using Transformers. Predicts a plausible user prompt from one assistant response. | |
| **Preview, supervised step 500** 路 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** 路 [Qwen2.5-0.5B-Instruct](https://huggingface.co/Qwen/Qwen2.5-0.5B-Instruct) 路 [UltraChat 200k](https://huggingface.co/datasets/HuggingFaceH4/ultrachat_200k) 路 [MiniLM sentence encoder used for evaluation](https://huggingface.co/sentence-transformers/all-MiniLM-L6-v2) 路 [Hu et al., LoRA](https://arxiv.org/abs/2106.09685) 路 [Zhang, Morris & Shmatikov, Output2Prompt (EMNLP 2024)](https://aclanthology.org/2024.emnlp-main.819/). | |