| --- |
| license: apache-2.0 |
| base_model: vllm-sr/d3-mini |
| base_model_relation: finetune |
| pipeline_tag: zero-shot-classification |
| library_name: transformers |
| language: |
| - en |
| tags: |
| - zero-shot-classification |
| - multi-label-classification |
| - decision-model |
| - classification |
| - system-one |
| - safetensors |
| datasets: |
| - google-research-datasets/go_emotions |
| - gfissore/arxiv-abstracts-2021 |
| - google/civil_comments |
| - coastalcph/lex_glue |
| - fancyzhx/dbpedia_14 |
| - TimSchopf/medical_abstracts |
| - AmazonScience/massive |
| - ucberkeley-dlab/measuring-hate-speech |
| - cardiffnlp/tweet_topic_multi |
| - google/boolq |
| - tau/commonsense_qa |
| - allenai/openbookqa |
| - allenai/qasc |
| - stanfordnlp/snli |
| --- |
| |
| # d3-mini-multi |
|
|
| **[d3-mini](https://huggingface.co/vllm-sr/d3-mini) with a fourth question type. `multi` answers "which of these |
| options apply?" with every option that applies, from zero to many, in one forward pass.** |
|
|
| | Type | Answers | Returns | |
| |---|---|---| |
| | `choice` | Which single option fits best? | One option | |
| | `score` | Where does this sit on an ordered scale? | One level | |
| | `noul` | Is this statement true? | One probability | |
| | **`multi`** | **Which options apply?** | **Every option that applies, a calibrated score for each, and the cutoff** | |
|
|
| `multi` adds one option to the question, the **cutoff**. The model scores the cutoff in the same pass and the |
| same softmax as the labels, and keeps every label that outscores it. Because the cutoff is scored from the |
| input, the model sets a new threshold for every input. The default cutoff works without tuning. An offset is there when a task needs it. |
|
|
| - **Code and proxy:** [github.com/akta-pro/d3-multi](https://github.com/akta-pro/d3-multi) |
| - **Paper:** [The Cutoff Is an Option](https://akta.pro/blog/the-cutoff-is-an-option-single-pass-multi-label-classification-with-adaptive-thresholding) |
|
|
| ## Results |
|
|
| Micro-F1 on held-out test data, with no threshold tuning on either side. "1 pass" means one forward pass per |
| input. |
|
|
| | Test set | Labels | Base: one `noul` per label (p > 0.5) | Base: `choice` top-1 | **d3-mini-multi (1 pass)** | |
| |---|---|---|---|---| |
| | arXiv categories | 155 | 0.29 | 0.65 | **0.67** | |
| | Civil Comments (toxicity attributes) | 7 | 0.59 | 0.41 | **0.78** | |
| | DBpedia-14 | 14 | 0.90 | 0.97 | **0.98** | |
| | ECtHR-A (articles violated) | 10 | 0.66 | 0.69 | **0.75** | |
| | ECtHR-B (articles alleged) | 10 | 0.69 | 0.72 | **0.80** | |
| | GoEmotions | 28 | 0.40 | **0.56** | **0.56** | |
| | **UNFAIR-ToS (never seen in training)** | 8 | 0.60 | 0.19 | **0.66** | |
| | **TweetTopic (never seen in training)** | 19 | 0.63 | **0.68** | **0.68** | |
| | **Hate-speech targets (never seen in training)** | 8 | 0.69 | **0.76** | 0.72 | |
|
|
| `multi` beats per-label `noul` on all nine test sets, including the three domains that are not in the training |
| data. Against `choice` top-1, it wins on six, ties on two, and trails on hate-speech targets, where most comments |
| target a single group. |
|
|
| **`choice`, `score` and `noul` keep their accuracy.** We ran the base and fine-tuned models on the same 10,763 requests |
| (29,053 questions) from 42 public benchmarks of the Decision Index. The public index is 55.10 for d3-mini-multi |
| and 55.05 for the base model: a change of +0.01 points, with a standard error of 0.22. |
|
|
| The tuned-baseline comparison, cost and calibration are in the |
| [GitHub README](https://github.com/akta-pro/d3-multi#results). |
|
|
| ## Quickstart |
|
|
| Serve the model with vLLM, then put the `multi` proxy from the GitHub repository in front of it: |
|
|
| ```bash |
| git clone https://github.com/akta-pro/d3-multi && cd d3-multi |
| pip install -r requirements.txt |
| vllm serve akta-pro/d3-mini-multi --served-model-name d3-mini-multi --port 8100 --no-enable-prefix-caching |
| python -m d3multi.serve --upstream http://localhost:8100 --port 8200 |
| ``` |
|
|
| Ask a `multi` question: |
|
|
| ```bash |
| curl -s localhost:8200/v1/systemone -H 'content-type: application/json' -d '{ |
| "model": "d3-mini-multi", |
| "state": "The order arrived damaged yesterday. The customer was charged twice and wants a replacement today.", |
| "questions": { |
| "teams": { |
| "type": "multi", |
| "instructions": "Which teams need to act on this request?", |
| "criteria": { |
| "returns": "Refunds, replacements and damaged deliveries", |
| "billing": "Payments, invoices and charges", |
| "technical": "Product setup and faults", |
| "shipping": "Delivery tracking and couriers" |
| } |
| } |
| } |
| }' |
| ``` |
|
|
| The answer (scores rounded to four decimals): |
|
|
| ```json |
| { |
| "teams": { |
| "type": "multi", |
| "selected": ["returns", "billing"], |
| "scores": { |
| "returns": 0.9867, |
| "billing": 0.9108, |
| "shipping": 0.2438, |
| "technical": 0.0494 |
| }, |
| "cutoff": 0.0116 |
| } |
| } |
| ``` |
|
|
| `choice`, `score` and `noul` questions pass through the proxy unchanged, and they can share a request with |
| `multi` questions. The request format, the `offset` setting and the client-side decoder are in the |
| [GitHub repository](https://github.com/akta-pro/d3-multi). |
|
|
| ## Limits |
|
|
| - **Unseen domains.** On UNFAIR-ToS, untuned `multi` reaches recall 0.94 but precision 0.51: it tags too many |
| sentences on a set where 89% of inputs have no label. On hate-speech targets, where most comments target one |
| group, `choice` top-1 is ahead (0.76 against 0.72). |
| - **254 options per question.** The readout scores 255 codes, and the cutoff takes one. |
| - **Near-zero margins.** Options within a hair of the cutoff can flip with batch shape or numeric precision. |
|
|
| <sub>All test sets are English text. We did not test other languages, or `multi` questions about images or video.</sub> |
|
|
| ## Files |
|
|
| The weights are BF16. The readout, tokenizer, configuration and d3 runtime files are copied unchanged from |
| [vllm-sr/d3-mini](https://huggingface.co/vllm-sr/d3-mini) at revision 97dd054. Only the language-model weights |
| differ. `SHA256SUMS` lists the hash of each weight file. |
|
|
| ## License |
|
|
| Apache-2.0 ([LICENSE](LICENSE), [NOTICE](NOTICE)). d3-mini is Apache-2.0 (Copyright 2026 vLLM Semantic Router |
| Team), built on [Qwen/Qwen3.5-4B](https://huggingface.co/Qwen/Qwen3.5-4B) (Apache-2.0). |
|
|
| ## Citation |
|
|
| ```bibtex |
| @misc{d3_multi_2026, |
| title = {The Cutoff Is an Option: Single-Pass Multi-Label Classification |
| with Adaptive Thresholding}, |
| author = {Balot, Harsh and {akta.pro}}, |
| year = {2026}, |
| howpublished = {\url{https://akta.pro/blog/the-cutoff-is-an-option-single-pass-multi-label-classification-with-adaptive-thresholding}}, |
| note = {Code: \url{https://github.com/akta-pro/d3-multi}} |
| } |
| ``` |
|
|
