| --- |
| license: mit |
| library_name: other |
| pipeline_tag: text-classification |
| tags: |
| - text-classification |
| - intent-detection |
| - router |
| - multilingual |
| - cpu |
| language: |
| - pt |
| - en |
| --- |
| |
| # Murph |
|
|
| Murph is the small request classifier built into [openTARS](https://project-opentars.github.io/OpenTARS/). For each request it decides the **kind of task**, so the app can pick the right model and tools before any language model runs. |
|
|
| It is not a language model. It is a softmax logistic regression over word pieces (2 to 5 letters), words and word pairs, with TF-IDF weights. It runs in plain Python on the CPU in about 0.2 ms per request, uses no GPU, and the whole model is one JSON file of about 4 MB. |
|
|
| ## Classes |
|
|
| | Class | Meaning | |
| |---|---| |
| | `visao` | needs to look at the screen ("what is on this screen?") | |
| | `codigo` | writing or fixing code | |
| | `tecnico` | technical or system questions | |
| | `acao` | do something on the computer (open a program, create a folder) | |
| | `busca` | search the web | |
| | `simples` | greetings and simple chat | |
| | `geral` | everything else | |
|
|
| ## Files |
|
|
| - `murph.json`: the trained model (SHA-256 `f1752425196a67c4baa28de5f6564a8dbccb6d41d12f44a0d0459a888c078f45`). |
| - `tars_murph.py`: the loader and classifier. It uses only the Python standard library. |
|
|
| ## Use |
|
|
| ```python |
| import tars_murph |
| |
| murph = tars_murph.Murph.carregar("murph.json") |
| tarefa, confianca = murph.classificar("abre a calculadora") |
| print(tarefa, round(confianca, 2)) # acao 0.98 |
| ``` |
|
|
| The model also stores a confidence threshold and a coverage minimum, so the app can fall back to another method when the request is unlike anything it saw in training. |
|
|
| ## Training and results |
|
|
| - Trained on 3,034 request phrases in five languages: examples from the openTARS project plus phrases written by a larger model. Phrases similar to the evaluation phrases were removed from training. |
| - Cross-validation accuracy stored in the file: 88.9%. |
| - openTARS evaluation sets, measured on 18 October 2026: 170 of 175 across languages and 85 of 86 on the check set. The evaluation phrases were seen during development, so treat these as a sanity check, not as a benchmark. |
|
|
| ## Limits |
|
|
| - Five languages only, and short requests work best. |
| - It picks one class per request. A request that mixes two tasks gets the more likely one. |
| - The class names are Portuguese because they come from the app code. |
|
|
| License: MIT. |
|
|
