llama-system1

Llama.cpp prompt template hack to perform fast one-shot single output token classification.

Usage

As a fully qualified /v1/systemone frontend:

$ npx llama-system1 8080 https://127.0.0.1:9931
Listening on http://127.0.0.1:8080/v1/systemone -> http://127.0.0.1:9931

Programmatic api

import { infer } from 'llama-system1'

const input = {
  state: 'There are 5 bananas on the table.',
  questions: {
    easy: {
      instructions: 'Were fruits mentioned?',
      type: 'noul'
   }
  }
}

const output = await infer(null, input)

console.log(output)

Used on qwen3.5:9b produces output:

{
  answers: { easy: { type: 'noul', noul: 0.9999997270783773 } },
  usage: { input_tokens: 96, output_tokens: 2 }
}

Performance

Measured locally against the 231 public JevBench decisions

| Metric | Result | | --- | ---: | | Accuracy | 69.70% (161 / 231) | | Brier score | 0.5656 | | Expected calibration error | 0.2599 | | Schema validity | 100% (231 / 231) | | Median latency | 48.1 ms | | p95 latency | 297.3 ms | | Ordinal MAE | 0.3113 |

License

ISC

All wrongs reversed - 2026 - Tony Ivanov