Feature Extraction
Transformers
Safetensors
English
modernbert
decision-model
intent-routing
text-embeddings-inference
Instructions to use aneuwirth/tars with libraries, inference providers, notebooks, and local apps. Follow these links to get started.
- Libraries
- Transformers
How to use aneuwirth/tars with Transformers:
# Use a pipeline as a high-level helper from transformers import pipeline pipe = pipeline("feature-extraction", model="aneuwirth/tars")# pip install -U transformers accelerate # Load model directly from transformers import AutoTokenizer, AutoModel tokenizer = AutoTokenizer.from_pretrained("aneuwirth/tars") model = AutoModel.from_pretrained("aneuwirth/tars", device_map="auto") - Notebooks
- Google Colab
- Kaggle
|
Download README.md from aneuwirth/tars: direct link, hf CLI and curl.
- Browser
- Download file 5.6 kB
-
https://huggingface.co/aneuwirth/tars/resolve/main/README.md
- Command line
-
hf download hf://aneuwirth/tars/README.md
-
curl -L -o README.md https://huggingface.co/aneuwirth/tars/resolve/main/README.md
5.6 kB
| license: apache-2.0 | |
| base_model: wfzyx/von | |
| library_name: transformers | |
| language: en | |
| tags: | |
| - decision-model | |
| - intent-routing | |
| - modernbert | |
| # Model card: Tars 0.5 | |
| ## Summary | |
| Tars is a 395M-parameter decision model that takes a spoken command and decides what to do with it: where one | |
| request ends and the next begins, which part of the assistant handles each request, and which skill to run. It | |
| does not generate text. It picks one option from a list you supply and gives a probability for it. | |
| Tars 0.5 is trained only on public, synthetic data: labels for the made-up example catalog in the code repository. | |
| It is meant as a starting point. Fine-tune it on your own catalog (the Tars Mac app's "Start from Tars" does this), | |
| since it only knows the example catalog's skills well. | |
| | | | | |
| |---|---| | |
| | Base model | [Von](https://github.com/wfzyx/von) 1.2 by Victor Hugo (Apache-2.0), a ModernBERT option-marker decision model | | |
| | Encoder | [ModernBERT](https://github.com/AnswerDotAI/ModernBERT) (Answer.AI and LightOn, Apache-2.0) | | |
| | Fine-tuning | Full fine-tune (encoder and scoring head) | | |
| | Size | 395M parameters, about 2.9 GB on disk with tokenizer and calibration | | |
| | License | Apache-2.0 | | |
| | Code, labels and benchmark | https://github.com/andrewneuwirth/Tars | | |
| ## Intended use | |
| - Deciding which skill, tool or app should handle a short command, given a catalog of skills you write. | |
| - Splitting a sentence that asks for several things into steps. | |
| - A starting point for fine-tuning on your own catalog. | |
| - Running locally on a laptop as a fast first pass. Act on Tars's decision above a probability threshold and fall | |
| back to your own rules or a larger model below it. | |
| It is not a chat model, cannot answer questions, and cannot take actions by itself. | |
| ## Training data | |
| - **Synthetic labels for the public example catalog** (`catalogs/example.json`: made-up apps such as Tunebox and | |
| Inkpad, 10 families, 31 skills), written by Anthropic's Claude through `pipeline/generate.py`: messy single-skill | |
| phrasings in 13 speaking styles (terse, casual, speech-to-text errors with no punctuation, rambling, and more), | |
| 2 to 5 step chained commands, near-miss pairs between similar skills, and not-a-command chatter (`none`). Every | |
| label was checked against the catalog and deduplicated. 8,432 labeled sentences, published in the code | |
| repository as `labels/example.jsonl`. Speech recognition noise is added in the build step. | |
| - **General-reasoning replay**: 13,000 rows from Von's own synthetic decision generator, mixed in so fine-tuning | |
| does not erase what the base model already did. | |
| - No personal data: no logged commands, no personal catalog. | |
| - 76,542 training rows (63,542 split, route and skill rows plus 13,000 replay), 3,124 optimizer steps (batch 8 x 4, | |
| learning rate 2e-5), 151 minutes on an Apple M5 Pro. The checkpoint is the one from step 3,000, picked on | |
| validation plus general replay, never on the benchmark. | |
| ## Evaluation | |
| The public benchmark (`benchmark/public/benchmark.jsonl` in the code repository) has 477 hard multi-step commands | |
| for the example catalog, 2 to 5 steps each. It was generated separately from the training labels, in three speaking | |
| styles never used in training (formal dictation, gamer talk, a non-technical parent). No training sentence has 80% | |
| or more word overlap with a benchmark sentence. A sentence counts only if every step's words, family and skill are | |
| right. | |
| | Model | End to end | Step count right | Split exact | Families right | Skills right | Median latency | | |
| |---|---|---|---|---|---|---| | |
| | Von 1.2 (base, no fine-tune) | 0.0% | 1.7% | 0.4% | 0.0% | 0.0% | 791 ms | | |
| | **Tars 0.5** | **69.2%** | **77.8%** | **71.3%** | **74.8%** | **74.8%** | **280 ms** | | |
| Latency is the whole plan (every split, route and skill question), in-process on an Apple M5 Pro (MPS). | |
| Each question alone, given the correct steps (`pipeline/evaluate.py`, same 477 sentences): split 44.2% → 98.6%, | |
| route 29.0% → 98.8%, skill 74.7% → 100.0%, whole sentence 0.0% → 90.1%. General reasoning on 600 of Von's replay | |
| rows went from 84.3% to 91.2%. | |
| Tars 0.4's numbers (65.0% end to end) came from a private benchmark for a personal catalog and can't be compared | |
| directly with these. | |
| ## Limits | |
| - It knows the example catalog's skills. Your own catalog needs a fine-tune; this model is the starting point. | |
| - Most errors are extra splits: a filler word or aside at the start or middle ("yo", "please", "cause the raid | |
| starts then") becomes its own `none` step. 92 of the 147 failed benchmark sentences have too many steps. | |
| - Labels written by an LLM carry that LLM's mistakes and its idea of how people talk. A few benchmark labels are | |
| debatable too. | |
| - English only. Tested on short commands; long dictation is out of scope. | |
| - Probabilities come from Von 1.2's calibration and are not refit after fine-tuning. On the benchmark they are close | |
| to 1.0 for almost every step, including steps in sentences it gets wrong, so they are a weak signal. Check them on | |
| your own data before choosing a threshold. | |
| ## Use | |
| ```python | |
| from tars import Catalog, Planner # https://github.com/andrewneuwirth/Tars | |
| planner = Planner(Catalog.load("catalogs/example.json"), "path/to/this/folder") | |
| planner.plan("open tunebox and play some jazz then mute") | |
| ``` | |
| The Tars Mac app downloads this checkpoint to fine-tune from it ("Start from Tars 0.5" in Settings). | |
| ## Attribution | |
| Tars is a fine-tune of [Von](https://github.com/wfzyx/von) 1.2 by Victor Hugo, Apache-2.0. See NOTICE. The training | |
| labels and benchmark were written by Anthropic's Claude. | |