Text Classification
Transformers
Safetensors
ONNX
multilingual
modernbert
routing
llm-routing
prompt-classification
complexity-estimation
text-embeddings-inference
Instructions to use scombat/s1d1-router with libraries, inference providers, notebooks, and local apps. Follow these links to get started.
- Libraries
- Transformers
How to use scombat/s1d1-router with Transformers:
# Use a pipeline as a high-level helper from transformers import pipeline pipe = pipeline("text-classification", model="scombat/s1d1-router")# pip install -U transformers accelerate # Load model directly from transformers import AutoTokenizer, AutoModelForSequenceClassification tokenizer = AutoTokenizer.from_pretrained("scombat/s1d1-router") model = AutoModelForSequenceClassification.from_pretrained("scombat/s1d1-router", device_map="auto") - Notebooks
- Google Colab
- Kaggle
Download integrations/policy.example.yaml from scombat/s1d1-router: direct link, hf CLI and curl.
- Browser
- Download file 3.29 kB
-
https://huggingface.co/scombat/s1d1-router/resolve/main/integrations/policy.example.yaml
- Command line
-
hf download hf://scombat/s1d1-router/integrations/policy.example.yaml
-
curl -L -o policy.example.yaml https://huggingface.co/scombat/s1d1-router/resolve/main/integrations/policy.example.yaml
3.29 kB
| # Example policy for s1d1-router: which model and which reasoning effort for each request. | |
| # Rules are tried in order; the first one that matches wins. Thresholds sit halfway between the tier scores | |
| # (trivial 0.01 | low 0.25 | medium 0.50 | high 0.75; the score itself goes up to 1.00), so 0.375 / 0.625 separate | |
| # the tiers and 0.875 isolates the top of the scale. | |
| # START HERE, THEN CALIBRATE ON YOUR OWN TRAFFIC: these numbers are a reasonable default, not a validated optimum. | |
| version: 1 | |
| effort_scale: [low, medium, high, xhigh, max] # from least to most reasoning | |
| # Used when no rule matches, and when the router is unavailable or too slow (set your own timeout, ~150 ms). | |
| default: {tier: mid, effort: medium} | |
| # A prompt that starts with one of these strings is forced to that tier (the user always wins). | |
| overrides: | |
| "!small": small | |
| "!mid": mid | |
| "!large": large | |
| "!frontier": frontier | |
| # Within one session: never go down once escalated (a model switch breaks the prompt cache and the | |
| # reasoning blocks of the previous turns), and switch at most N times. | |
| session: | |
| escalate_only: true | |
| max_switches: 6 | |
| rules: | |
| # 1. The top of the scale: the top model, high effort. | |
| - name: very-hard | |
| when: {complexity_min: 0.875} | |
| tier: frontier | |
| effort: xhigh | |
| # 2. Hard prompts. Code, maths, reasoning and planning get more thinking than the rest. | |
| - name: hard-technical | |
| when: {complexity_min: 0.625, task_in: [coding_debugging, math, reasoning, planning, data_analysis]} | |
| tier: large | |
| effort: high | |
| - name: hard-other | |
| when: {complexity_min: 0.625} | |
| tier: large | |
| effort: medium | |
| # 3. Anything that acts on tools (files, shell, browser, external data) never goes to the smallest tier: | |
| # small models make more tool-call mistakes, and a failed loop costs more than a bigger model. | |
| - name: agentic | |
| when: {needs_tools: true, complexity_max: 0.625} | |
| tier: mid | |
| effort: medium | |
| - name: agentic-task | |
| when: {task_in: [action_execution, data_analysis], complexity_max: 0.625} | |
| tier: mid | |
| effort: medium | |
| # 4. Medium prompts. | |
| - name: medium-reasoning | |
| when: {complexity_min: 0.375, task_in: [math, reasoning, planning]} | |
| tier: large | |
| effort: medium | |
| - name: medium-code | |
| when: {complexity_min: 0.375, task_in: [coding_debugging]} | |
| tier: mid | |
| effort: medium | |
| - name: medium-other | |
| when: {complexity_min: 0.375} | |
| tier: mid | |
| effort: low | |
| # 5. Trivial and low prompts without tools: the smallest model, minimal reasoning. | |
| - name: easy | |
| when: {complexity_max: 0.375, needs_tools: false} | |
| tier: small | |
| effort: low | |
| # One profile per harness: how the tiers map to its model ids, and which efforts it accepts. | |
| profiles: | |
| claude_code: | |
| models: {small: claude-haiku-5-5, mid: claude-sonnet-5-5, large: claude-opus-5-5, frontier: claude-fable-5-1} | |
| efforts: [low, medium, high, xhigh, max] # request body field output_config.effort | |
| codex: | |
| models: {small: "<your-small-model>", mid: "<your-mid-model>", large: "<your-large-model>", frontier: "<your-top-model>"} | |
| efforts: [low, medium, high, xhigh] # config key model_reasoning_effort | |
| generic: | |
| models: {small: local-small, mid: cloud-mid, large: cloud-large, frontier: cloud-top} | |
| efforts: [low, medium, high] | |