98 lines
3.3 KiB
Go
98 lines
3.3 KiB
Go
package judgegate
|
|
|
|
import (
|
|
"bytes"
|
|
"context"
|
|
"encoding/json"
|
|
"flag"
|
|
"fmt"
|
|
"os"
|
|
"time"
|
|
)
|
|
|
|
// EvalSpec is what a guard supplies to measure its judge.
|
|
type EvalSpec[T any] struct {
|
|
// Name is the guard's name, used in messages.
|
|
Name string
|
|
// KeyEnv is the environment variable holding the judge's API key, when the
|
|
// endpoint needs one.
|
|
KeyEnv string
|
|
Threshold Threshold
|
|
// Load reads the labelled cases from path.
|
|
Load func(path string) ([]T, error)
|
|
// Judge asks the judge about one case and grades its answer against the
|
|
// case's label. On an error it still returns the case's id and label.
|
|
Judge func(ctx context.Context, c Client, tc T) (Graded, error)
|
|
}
|
|
|
|
// RunEval is a guard's cmd/eval: it measures the judge named on the command
|
|
// line against the labelled cases and writes the record the guard reads at start.
|
|
func RunEval[T any](spec EvalSpec[T]) error {
|
|
api := flag.String("api", APISystemOne, "judge backend: systemone or openai-decisions")
|
|
base := flag.String("base", "", "judge base URL; required for systemone, defaults for openai-decisions")
|
|
model := flag.String("model", "", "judge model id; defaults for openai-decisions")
|
|
cases := flag.String("cases", "eval/cases.json", "labelled cases")
|
|
out := flag.String("out", "eval_result.json", "where to write the result")
|
|
answers := flag.String("answers", "eval/answers.json", "where to write every case's answer for review")
|
|
flag.Parse()
|
|
key := os.Getenv(spec.KeyEnv)
|
|
if *api == APIOpenAIDecisions {
|
|
key = os.Getenv("OPENAI_API_KEY")
|
|
}
|
|
client, err := EvalClient(*api, *base, *model, key)
|
|
if err != nil {
|
|
return err
|
|
}
|
|
loaded, err := spec.Load(*cases)
|
|
if err != nil {
|
|
return err
|
|
}
|
|
measured := Measure(context.Background(), loaded, func(ctx context.Context, tc T) (Graded, error) {
|
|
return spec.Judge(ctx, client, tc)
|
|
})
|
|
rec := NewRecord(*model, time.Now(), measured, spec.Threshold)
|
|
fmt.Printf("%s: %d cases answered, %d judge errors, disagreeing with the label %v\n", spec.Name, measured.Cases, measured.Errors, measured.Disagreed)
|
|
if measured.FirstError != "" {
|
|
fmt.Println("first judge error:", measured.FirstError)
|
|
}
|
|
fmt.Printf("recall %.2f precision %.2f false flag rate %.2f accuracy %.2f\n",
|
|
rec.Judged.Recall, rec.Judged.Precision, rec.Judged.FalseFlagRate, rec.Judged.Accuracy)
|
|
data, err := json.MarshalIndent(rec, "", " ")
|
|
if err != nil {
|
|
return err
|
|
}
|
|
if err := os.WriteFile(*out, append(data, '\n'), 0o600); err != nil {
|
|
return err
|
|
}
|
|
review := struct {
|
|
Model string `json:"model"`
|
|
Date string `json:"date"`
|
|
Answers []Graded `json:"answers"`
|
|
}{rec.Model, rec.Date, measured.Answers}
|
|
data, err = json.MarshalIndent(review, "", " ")
|
|
if err != nil {
|
|
return err
|
|
}
|
|
if err := os.WriteFile(*answers, append(data, '\n'), 0o600); err != nil {
|
|
return err
|
|
}
|
|
if short := rec.Shortfalls(); len(short) > 0 {
|
|
fmt.Println("the judge stays advisory:", short)
|
|
} else {
|
|
fmt.Println("the judge passes its threshold")
|
|
}
|
|
return nil
|
|
}
|
|
|
|
// ReadJSON decodes data into v and rejects members v doesn't know.
|
|
func ReadJSON(data []byte, v any) error {
|
|
dec := json.NewDecoder(bytes.NewReader(data))
|
|
dec.DisallowUnknownFields()
|
|
return dec.Decode(v)
|
|
}
|
|
|
|
// ReadFile reads the cases file the operator named.
|
|
func ReadFile(path string) ([]byte, error) {
|
|
return os.ReadFile(path) //nolint:gosec // the operator names the cases file
|
|
}
|