Multilingual Prompt Task & Complexity Classifier (base, 308M)

Classifies a user prompt sent to an LLM, in English and 100+ other languages:

• task type (11, the taxonomy of NVIDIA's prompt classifier): open QA, closed QA, summarization, text generation, code generation, chatbot, classification, rewrite, brainstorming, extraction, other;
• complexity scores from 0 to 1: creativity, reasoning, contextual knowledge, domain knowledge, constraints, and overall complexity.

Typical uses: routing prompts to a cheaper or stronger model, analysing chat logs, balancing instruction-tuning data. Built on mmBERT-base, Apache-2.0, ONNX for CPU and the browser (transformers.js) included. A smaller, faster version is available as Horizon-Labs/multilingual-prompt-task-complexity-small.

• 17 sigmoid outputs: task_* (take the highest as the task type) and the six dimensions. onnx/model_quantized.onnx (int8 embeddings, 641 MB) gives the same 0.5 decisions as fp32 on all outputs for 97.8% of 400 test prompts.
• The dimension scores are relative: most prompts score low (training averages: complexity 0.20, reasoning 0.21, creativity 0.29). Compare prompts with each other, or pick routing thresholds from your own traffic.

Usage

from transformers import pipeline

clf = pipeline("text-classification", model="Horizon-Labs/multilingual-prompt-task-complexity-base", top_k=None)
s = {d["label"]: d["score"] for d in clf("Écris une fonction Python qui vérifie si une chaîne est un palindrome.")[0]}
task = max((k for k in s if k.startswith("task_")), key=s.get)
print(task, round(s["complexity"], 2), round(s["reasoning"], 2))

Evaluation

There is no public human-labelled multilingual test set for this taxonomy. We held out 4,200 real user prompts from WildChat-1M (200 in each of 21 languages, never used for training) and labelled them with an independent judge, google/gemma-4-31b-it, using the same questions as our teacher. Task accuracy and macro-F1 compare the top task; the dimensions use Spearman rank correlation.

• The table shows the released checkpoint. Means over 2 training seeds (small: task accuracy .793 / .790, dimension Spearman .716 / .716; base: .811 / .813, .741 / .740).
• Read the comparison with NVIDIA's classifier with care: it was trained on human annotations with its own guidelines, so part of its disagreement with our judge is about definitions, not errors; on English prompts the two agree on the task 66% of the time. Outside English its agreement drops further, which is the gap this model fills.
• Caveat: the judge and our teacher are both LLMs and agree with each other closely (task 86%), which favours this model on this test.

Training

• Prompts: 172,690 real user prompts in 133 languages: first user messages from WildChat-1M (ODC-BY; only the prompt text was used, no metadata; conversations flagged toxic excluded), prompts from OASST2 and Aya (Apache-2.0). At most 3,000 per language (English 20,000, Chinese 8,000).
• Labels: Qwen3.8-27B (Apache-2.0) answered one multiple-choice question (task) and six 0-9 ratings per prompt; the model learns the answer probabilities (task) and expected ratings / 9 (dimensions) as soft targets. Checkpoint chosen on 8,589 held-out prompts.
• mmBERT-base, max length 512 tokens. Code: code/ in this repository.

Limitations

• Labels come from an LLM, not people, and complexity ratings are subjective. Use the scores for ranking and routing, and validate thresholds on your own traffic.
• Only the first 512 tokens of a prompt are read; very long prompts with instructions at the end may be misclassified.
• Not a safety classifier: for harmful or injected prompts see our content-safety-guard and prompt-injection-guard.