nadirclaw-distilbert

A 3-class prompt complexity classifier for NadirClaw, an open-source LLM router. Given a prompt, it predicts which routing tier the request belongs to so the router can send simple prompts to cheap models and hard prompts to capable ones.

Classes

Label	Tier	Examples
0	simple	short questions, lookups, formatting, file reads
1	medium (→ mid in NadirClaw)	focused single-function edits, single-file debugging, small refactors
2	complex	architecture, multi-file refactors, agentic loops, reasoning-heavy prompts

Model details

• Base: distilbert-base-uncased, native sequence-classification head (num_labels=3)
• Inference: CPU, ~30ms per prompt, max_length=256
• Artifact size: ~256MB
• Training data: NadirClaw + Horizen prototype prompts + eval prompts (~700 samples)

Usage

from transformers import DistilBertTokenizer, DistilBertForSequenceClassification
import torch

repo = "nadirclaw/nadirclaw-distilbert"
tok = DistilBertTokenizer.from_pretrained(repo)
model = DistilBertForSequenceClassification.from_pretrained(repo)
model.eval()

inputs = tok("design a distributed rate limiter", return_tensors="pt",
             truncation=True, max_length=256)
with torch.no_grad():
    probs = torch.softmax(model(**inputs).logits[0], dim=-1)
tier = ["simple", "medium", "complex"][int(probs.argmax())]
print(tier, float(probs.max()))

In NadirClaw, just set:

NADIRCLAW_COMPLEXITY_ANALYZER=distilbert

NadirClaw downloads and caches this model automatically on first use.

Calibration

The classifier leans toward over-serving rather than under-serving: when a simple prediction has confidence < 0.70, it escalates to medium or complex. It is cheaper to over-serve a simple prompt than to under-serve a complex one.

License

Apache-2.0, matching the base model. NadirClaw itself is MIT.