mirror of
https://github.com/lukaszraczylo/llm-testbench.git
synced 2026-10-12 00:45:19 +00:00
lukaszraczylo
1d51d4cbbd
Framework runs a category/subcategory test catalog against multiple LLMs through an OpenAI-compatible endpoint. Deterministic evaluators score each response. Reports render as table, markdown, or JSON.
97 lines
3.0 KiB
Go
97 lines
3.0 KiB
Go
package eval
|
|
|
|
import (
|
|
"context"
|
|
"testing"
|
|
)
|
|
|
|
func TestExtractLastNumber(t *testing.T) {
|
|
tests := []struct {
|
|
name string
|
|
response string
|
|
want float64
|
|
wantErr bool
|
|
}{
|
|
{"plain number", "0.8785", 0.8785, false},
|
|
{"trailing number after prose", "The cosine similarity is 0.8785", 0.8785, false},
|
|
{"last of multiple numbers", "step 1 then step 2 result is 42", 42, false},
|
|
{"negative number", "the delta is -3.5", -3.5, false},
|
|
{"no number", "no digits here", 0, true},
|
|
// S6 regressions.
|
|
{"unit qualifier after the answer is skipped", "24 bytes on a 64-bit system", 24, false},
|
|
{"leading-dot decimal", "= .9896", 0.9896, false},
|
|
{"answer on last line, reasoning with numbers on earlier lines", "step 1 uses 8 items\nstep 2 uses 16 items\nfinal answer: 42", 42, false},
|
|
{"last line has only a unit-qualified number, falls back to whole text", "the computation used 100 iterations\nresult on a 64-bit system", 100, false},
|
|
}
|
|
for _, tt := range tests {
|
|
t.Run(tt.name, func(t *testing.T) {
|
|
got, err := ExtractLastNumber[float64](tt.response)
|
|
if (err != nil) != tt.wantErr {
|
|
t.Fatalf("ExtractLastNumber(%q) error = %v, wantErr %v", tt.response, err, tt.wantErr)
|
|
}
|
|
if err == nil && got != tt.want {
|
|
t.Errorf("ExtractLastNumber(%q) = %v, want %v", tt.response, got, tt.want)
|
|
}
|
|
})
|
|
}
|
|
}
|
|
|
|
func TestNumeric_Float(t *testing.T) {
|
|
e := Numeric(ExtractLastNumber[float64], 0.8785, 0.0005)
|
|
|
|
tests := []struct {
|
|
name string
|
|
response string
|
|
want float64
|
|
}{
|
|
{"exact", "0.8785", 1},
|
|
{"within tolerance above", "0.8789", 1},
|
|
{"within tolerance below", "0.8781", 1},
|
|
{"outside tolerance", "0.9000", 0},
|
|
{"unparsable", "no number here", 0},
|
|
}
|
|
for _, tt := range tests {
|
|
t.Run(tt.name, func(t *testing.T) {
|
|
got := e.Evaluate(context.Background(), tt.response)
|
|
if got.Value != tt.want {
|
|
t.Errorf("Evaluate(%q) = %v, want %v (detail: %s)", tt.response, got.Value, tt.want, got.Detail)
|
|
}
|
|
})
|
|
}
|
|
}
|
|
|
|
func TestNumeric_IntExact(t *testing.T) {
|
|
e := Numeric(ExtractLastNumber[int], 32, 0)
|
|
|
|
tests := []struct {
|
|
name string
|
|
response string
|
|
want float64
|
|
}{
|
|
{"exact match", "32", 1},
|
|
{"off by one", "33", 0},
|
|
{"exact match with prose", "The sizeof value is 32 bytes", 1},
|
|
}
|
|
for _, tt := range tests {
|
|
t.Run(tt.name, func(t *testing.T) {
|
|
got := e.Evaluate(context.Background(), tt.response)
|
|
if got.Value != tt.want {
|
|
t.Errorf("Evaluate(%q) = %v, want %v (detail: %s)", tt.response, got.Value, tt.want, got.Detail)
|
|
}
|
|
})
|
|
}
|
|
}
|
|
|
|
func TestNumeric_UnsignedNoUnderflow(t *testing.T) {
|
|
// Regression: unsigned subtraction must not wrap when got < want.
|
|
e := Numeric(ExtractLastNumber[uint], 10, 2)
|
|
got := e.Evaluate(context.Background(), "9")
|
|
if got.Value != 1 {
|
|
t.Errorf("Evaluate(9) with want=10 tol=2 = %v, want 1 (detail: %s)", got.Value, got.Detail)
|
|
}
|
|
got = e.Evaluate(context.Background(), "5")
|
|
if got.Value != 0 {
|
|
t.Errorf("Evaluate(5) with want=10 tol=2 = %v, want 0 (detail: %s)", got.Value, got.Detail)
|
|
}
|
|
}
|