496 lines
14 KiB
Rust
496 lines
14 KiB
Rust
//! Reproducible evaluations for any OpenAI-compatible model server.
|
|
//!
|
|
//! Suites are declarative YAML. Results are append-only JSON artifacts suitable
|
|
//! for CI, regression comparisons, and future publication to an eval registry.
|
|
|
|
use anyhow::{bail, Context, Result};
|
|
use serde::{Deserialize, Serialize};
|
|
use serde_json::{json, Value};
|
|
use std::fs;
|
|
use std::io::{BufRead, BufReader, Write};
|
|
use std::path::{Path, PathBuf};
|
|
use std::process::{Command, Stdio};
|
|
use std::time::{Instant, SystemTime, UNIX_EPOCH};
|
|
|
|
#[derive(Debug, Deserialize)]
|
|
struct Suite {
|
|
#[serde(rename = "apiVersion")]
|
|
api_version: String,
|
|
kind: String,
|
|
metadata: Metadata,
|
|
#[serde(default)]
|
|
defaults: Defaults,
|
|
cases: Vec<Case>,
|
|
}
|
|
|
|
#[derive(Debug, Deserialize)]
|
|
struct Metadata {
|
|
name: String,
|
|
#[serde(default)]
|
|
version: u32,
|
|
#[serde(default)]
|
|
description: String,
|
|
/// Part of the published suite manifest surface; parsed so a suite
|
|
/// carrying tags loads, not read by the runner itself.
|
|
#[serde(default)]
|
|
#[allow(dead_code)]
|
|
tags: Vec<String>,
|
|
}
|
|
|
|
#[derive(Debug, Default, Deserialize)]
|
|
struct Defaults {
|
|
#[serde(default = "default_max_tokens")]
|
|
max_tokens: u32,
|
|
#[serde(default = "default_temperature")]
|
|
temperature: f64,
|
|
#[serde(default = "default_repeat")]
|
|
repeat: u32,
|
|
#[serde(default)]
|
|
system: String,
|
|
}
|
|
|
|
fn default_max_tokens() -> u32 {
|
|
128
|
|
}
|
|
fn default_temperature() -> f64 {
|
|
0.0
|
|
}
|
|
fn default_repeat() -> u32 {
|
|
1
|
|
}
|
|
|
|
#[derive(Debug, Deserialize)]
|
|
struct Case {
|
|
id: String,
|
|
#[serde(default)]
|
|
category: String,
|
|
prompt: String,
|
|
#[serde(default)]
|
|
max_tokens: Option<u32>,
|
|
#[serde(default)]
|
|
assertions: Vec<Assertion>,
|
|
}
|
|
|
|
#[derive(Debug, Deserialize)]
|
|
#[serde(tag = "type", rename_all = "snake_case")]
|
|
enum Assertion {
|
|
Exact { value: String },
|
|
Contains { value: String },
|
|
ContainsAny { values: Vec<String> },
|
|
NotContains { value: String },
|
|
MaxWords { value: usize },
|
|
}
|
|
|
|
#[derive(Debug, Serialize)]
|
|
struct Artifact {
|
|
schema_version: u32,
|
|
run_id: String,
|
|
suite: String,
|
|
suite_version: u32,
|
|
model: String,
|
|
base_url: String,
|
|
started_unix_ms: u128,
|
|
summary: Summary,
|
|
samples: Vec<Sample>,
|
|
}
|
|
|
|
#[derive(Debug, Serialize)]
|
|
struct Summary {
|
|
passed: usize,
|
|
total: usize,
|
|
score: f64,
|
|
mean_ttft_ms: f64,
|
|
p50_ttft_ms: f64,
|
|
p95_ttft_ms: f64,
|
|
mean_prefill_tps: f64,
|
|
mean_decode_tps: f64,
|
|
}
|
|
|
|
#[derive(Debug, Serialize)]
|
|
struct Sample {
|
|
case_id: String,
|
|
category: String,
|
|
repetition: u32,
|
|
passed: bool,
|
|
assertion_results: Vec<bool>,
|
|
output: String,
|
|
reasoning: String,
|
|
prompt_tokens: u64,
|
|
completion_tokens: u64,
|
|
ttft_ms: f64,
|
|
total_ms: f64,
|
|
prefill_tps: f64,
|
|
decode_tps: f64,
|
|
}
|
|
|
|
struct Completion {
|
|
output: String,
|
|
reasoning: String,
|
|
prompt_tokens: u64,
|
|
completion_tokens: u64,
|
|
ttft_ms: f64,
|
|
total_ms: f64,
|
|
}
|
|
|
|
/// One line of `eval ls`. Suites themselves stay private — a caller has no
|
|
/// business reaching into cases and assertions — but the catalogue is the
|
|
/// useful part and is shared by the CLI and the MCP server.
|
|
#[derive(Debug, Clone)]
|
|
pub struct SuiteSummary {
|
|
pub name: String,
|
|
pub version: u32,
|
|
pub cases: usize,
|
|
pub description: String,
|
|
}
|
|
|
|
pub fn catalogue(root: &Path) -> Result<Vec<SuiteSummary>> {
|
|
Ok(load_all(root)?
|
|
.into_iter()
|
|
.map(|s| SuiteSummary {
|
|
name: s.metadata.name,
|
|
version: s.metadata.version,
|
|
cases: s.cases.len(),
|
|
description: s.metadata.description,
|
|
})
|
|
.collect())
|
|
}
|
|
|
|
pub fn list(root: &Path) -> Result<()> {
|
|
println!("{:<20} {:>4} {:>5} DESCRIPTION", "SUITE", "VER", "CASES");
|
|
for s in catalogue(root)? {
|
|
println!(
|
|
"{:<20} {:>4} {:>5} {}",
|
|
s.name, s.version, s.cases, s.description
|
|
);
|
|
}
|
|
Ok(())
|
|
}
|
|
|
|
pub fn run(
|
|
root: &Path,
|
|
name: &str,
|
|
base_url: &str,
|
|
model: &str,
|
|
repeat: Option<u32>,
|
|
) -> Result<()> {
|
|
let suite = load(root, name)?;
|
|
// `kuda/v1` is the pre-rename contract; still accepted so an older suite keeps running.
|
|
let known_contract = matches!(suite.api_version.as_str(), "lumbridge/v1" | "kuda/v1");
|
|
if !known_contract || suite.kind != "EvalSuite" {
|
|
bail!(
|
|
"unsupported eval contract: {}/{}",
|
|
suite.api_version,
|
|
suite.kind
|
|
);
|
|
}
|
|
let reps = repeat.unwrap_or(suite.defaults.repeat).max(1);
|
|
println!(
|
|
"eval '{}' v{} → {} ({})",
|
|
suite.metadata.name, suite.metadata.version, model, base_url
|
|
);
|
|
let started = now_ms();
|
|
let mut samples = Vec::new();
|
|
for case in &suite.cases {
|
|
for repetition in 1..=reps {
|
|
print!(" {:<24} [{}/{}] ", case.id, repetition, reps);
|
|
std::io::stdout().flush().ok();
|
|
let c = stream_completion(base_url, model, &suite.defaults, case)?;
|
|
let assertion_results: Vec<bool> = case
|
|
.assertions
|
|
.iter()
|
|
.map(|a| evaluate(a, &c.output))
|
|
.collect();
|
|
let passed = assertion_results.iter().all(|v| *v);
|
|
let decode_seconds = ((c.total_ms - c.ttft_ms) / 1000.0).max(0.001);
|
|
let prefill_seconds = (c.ttft_ms / 1000.0).max(0.001);
|
|
let sample = Sample {
|
|
case_id: case.id.clone(),
|
|
category: case.category.clone(),
|
|
repetition,
|
|
passed,
|
|
assertion_results,
|
|
output: c.output,
|
|
reasoning: c.reasoning,
|
|
prompt_tokens: c.prompt_tokens,
|
|
completion_tokens: c.completion_tokens,
|
|
ttft_ms: c.ttft_ms,
|
|
total_ms: c.total_ms,
|
|
prefill_tps: c.prompt_tokens as f64 / prefill_seconds,
|
|
decode_tps: c.completion_tokens as f64 / decode_seconds,
|
|
};
|
|
println!(
|
|
"{} ttft={:.0}ms decode={:.1}tok/s",
|
|
if passed { "PASS" } else { "FAIL" },
|
|
sample.ttft_ms,
|
|
sample.decode_tps
|
|
);
|
|
samples.push(sample);
|
|
}
|
|
}
|
|
let summary = summarize(&samples);
|
|
let run_id = format!("{}-{}-{}", started, slug(name), slug(model));
|
|
let artifact = Artifact {
|
|
schema_version: 1,
|
|
run_id: run_id.clone(),
|
|
suite: suite.metadata.name,
|
|
suite_version: suite.metadata.version,
|
|
model: model.to_string(),
|
|
base_url: base_url.to_string(),
|
|
started_unix_ms: started,
|
|
summary,
|
|
samples,
|
|
};
|
|
let out_dir = root.join("eval-results");
|
|
fs::create_dir_all(&out_dir)?;
|
|
let out = out_dir.join(format!("{run_id}.json"));
|
|
fs::write(&out, serde_json::to_string_pretty(&artifact)?)?;
|
|
println!(
|
|
"\nscore {:.1}% ({}/{}) · mean TTFT {:.0}ms · prefill≈{:.1}tok/s · decode {:.1}tok/s",
|
|
artifact.summary.score * 100.0,
|
|
artifact.summary.passed,
|
|
artifact.summary.total,
|
|
artifact.summary.mean_ttft_ms,
|
|
artifact.summary.mean_prefill_tps,
|
|
artifact.summary.mean_decode_tps
|
|
);
|
|
println!("result {}", out.display());
|
|
Ok(())
|
|
}
|
|
|
|
fn stream_completion(
|
|
base_url: &str,
|
|
model: &str,
|
|
defaults: &Defaults,
|
|
case: &Case,
|
|
) -> Result<Completion> {
|
|
let url = format!("{}/chat/completions", base_url.trim_end_matches('/'));
|
|
let mut messages = Vec::new();
|
|
if !defaults.system.is_empty() {
|
|
messages.push(json!({"role":"system","content":defaults.system}));
|
|
}
|
|
messages.push(json!({"role":"user","content":case.prompt}));
|
|
let request = json!({
|
|
"model": model, "messages": messages, "stream": true,
|
|
"stream_options": {"include_usage": true},
|
|
"max_tokens": case.max_tokens.unwrap_or(defaults.max_tokens),
|
|
"temperature": defaults.temperature,
|
|
"chat_template_kwargs": {"enable_thinking": false}
|
|
});
|
|
let mut child = Command::new("curl")
|
|
.args([
|
|
"-sS",
|
|
"-N",
|
|
"-X",
|
|
"POST",
|
|
&url,
|
|
"-H",
|
|
"Content-Type: application/json",
|
|
"--data-binary",
|
|
"@-",
|
|
])
|
|
.stdin(Stdio::piped())
|
|
.stdout(Stdio::piped())
|
|
.stderr(Stdio::piped())
|
|
.spawn()
|
|
.context("starting curl (required for eval HTTP streaming)")?;
|
|
child
|
|
.stdin
|
|
.take()
|
|
.unwrap()
|
|
.write_all(request.to_string().as_bytes())?;
|
|
let start = Instant::now();
|
|
let mut first = None;
|
|
let mut output = String::new();
|
|
let mut reasoning = String::new();
|
|
let mut prompt_tokens = 0;
|
|
let mut completion_tokens = 0;
|
|
for line in BufReader::new(child.stdout.take().unwrap()).lines() {
|
|
let line = line?;
|
|
let Some(data) = line.strip_prefix("data: ") else {
|
|
continue;
|
|
};
|
|
if data == "[DONE]" {
|
|
break;
|
|
}
|
|
let v: Value = serde_json::from_str(data).context("parsing streamed completion")?;
|
|
if let Some(usage) = v.get("usage") {
|
|
prompt_tokens = usage
|
|
.get("prompt_tokens")
|
|
.and_then(Value::as_u64)
|
|
.unwrap_or(prompt_tokens);
|
|
completion_tokens = usage
|
|
.get("completion_tokens")
|
|
.and_then(Value::as_u64)
|
|
.unwrap_or(completion_tokens);
|
|
}
|
|
let delta = &v["choices"][0]["delta"];
|
|
let content = delta.get("content").and_then(Value::as_str).unwrap_or("");
|
|
let thought = delta
|
|
.get("reasoning")
|
|
.or_else(|| delta.get("reasoning_content"))
|
|
.and_then(Value::as_str)
|
|
.unwrap_or("");
|
|
if first.is_none() && (!content.is_empty() || !thought.is_empty()) {
|
|
first = Some(start.elapsed());
|
|
}
|
|
output.push_str(content);
|
|
reasoning.push_str(thought);
|
|
}
|
|
let status = child.wait()?;
|
|
if !status.success() {
|
|
bail!("completion request failed with {status}");
|
|
}
|
|
let total = start.elapsed().as_secs_f64() * 1000.0;
|
|
Ok(Completion {
|
|
output,
|
|
reasoning,
|
|
prompt_tokens,
|
|
completion_tokens,
|
|
ttft_ms: first.map(|d| d.as_secs_f64() * 1000.0).unwrap_or(total),
|
|
total_ms: total,
|
|
})
|
|
}
|
|
|
|
fn evaluate(a: &Assertion, output: &str) -> bool {
|
|
let normalized = output.trim();
|
|
match a {
|
|
Assertion::Exact { value } => normalized.eq_ignore_ascii_case(value.trim()),
|
|
Assertion::Contains { value } => normalized.to_lowercase().contains(&value.to_lowercase()),
|
|
Assertion::ContainsAny { values } => values
|
|
.iter()
|
|
.any(|v| normalized.to_lowercase().contains(&v.to_lowercase())),
|
|
Assertion::NotContains { value } => {
|
|
!normalized.to_lowercase().contains(&value.to_lowercase())
|
|
}
|
|
Assertion::MaxWords { value } => normalized.split_whitespace().count() <= *value,
|
|
}
|
|
}
|
|
|
|
fn summarize(samples: &[Sample]) -> Summary {
|
|
let mut ttft: Vec<f64> = samples.iter().map(|s| s.ttft_ms).collect();
|
|
ttft.sort_by(f64::total_cmp);
|
|
let mean = |f: fn(&Sample) -> f64| {
|
|
if samples.is_empty() {
|
|
0.0
|
|
} else {
|
|
samples.iter().map(f).sum::<f64>() / samples.len() as f64
|
|
}
|
|
};
|
|
let percentile = |p: f64| {
|
|
if ttft.is_empty() {
|
|
0.0
|
|
} else {
|
|
ttft[((ttft.len() - 1) as f64 * p).round() as usize]
|
|
}
|
|
};
|
|
let passed = samples.iter().filter(|s| s.passed).count();
|
|
Summary {
|
|
passed,
|
|
total: samples.len(),
|
|
score: if samples.is_empty() {
|
|
0.0
|
|
} else {
|
|
passed as f64 / samples.len() as f64
|
|
},
|
|
mean_ttft_ms: mean(|s| s.ttft_ms),
|
|
p50_ttft_ms: percentile(0.50),
|
|
p95_ttft_ms: percentile(0.95),
|
|
mean_prefill_tps: mean(|s| s.prefill_tps),
|
|
mean_decode_tps: mean(|s| s.decode_tps),
|
|
}
|
|
}
|
|
|
|
fn load(root: &Path, name: &str) -> Result<Suite> {
|
|
load_all(root)?
|
|
.into_iter()
|
|
.find(|s| s.metadata.name == name)
|
|
.with_context(|| format!("no eval suite named '{name}' in {}/evals", root.display()))
|
|
}
|
|
|
|
fn load_all(root: &Path) -> Result<Vec<Suite>> {
|
|
let dir = root.join("evals");
|
|
let mut suites = Vec::new();
|
|
if !dir.exists() {
|
|
return Ok(suites);
|
|
}
|
|
for entry in fs::read_dir(&dir)? {
|
|
let path: PathBuf = entry?.path();
|
|
if path
|
|
.file_name()
|
|
.and_then(|n| n.to_str())
|
|
.map(|n| n.ends_with(".eval.yaml"))
|
|
.unwrap_or(false)
|
|
{
|
|
let suite: Suite = serde_yaml::from_str(&fs::read_to_string(&path)?)
|
|
.with_context(|| format!("parsing {}", path.display()))?;
|
|
suites.push(suite);
|
|
}
|
|
}
|
|
suites.sort_by(|a, b| a.metadata.name.cmp(&b.metadata.name));
|
|
Ok(suites)
|
|
}
|
|
|
|
fn now_ms() -> u128 {
|
|
SystemTime::now()
|
|
.duration_since(UNIX_EPOCH)
|
|
.unwrap_or_default()
|
|
.as_millis()
|
|
}
|
|
fn slug(s: &str) -> String {
|
|
s.chars()
|
|
.map(|c| {
|
|
if c.is_ascii_alphanumeric() {
|
|
c.to_ascii_lowercase()
|
|
} else {
|
|
'-'
|
|
}
|
|
})
|
|
.collect()
|
|
}
|
|
|
|
#[cfg(test)]
|
|
mod tests {
|
|
use super::*;
|
|
|
|
#[test]
|
|
fn assertions_are_case_insensitive_and_composable() {
|
|
let output = "Market risk remains, while company-specific risk falls.";
|
|
assert!(evaluate(
|
|
&Assertion::Contains {
|
|
value: "MARKET RISK".into()
|
|
},
|
|
output
|
|
));
|
|
assert!(evaluate(
|
|
&Assertion::ContainsAny {
|
|
values: vec!["idiosyncratic".into(), "company-specific".into()],
|
|
},
|
|
output
|
|
));
|
|
assert!(evaluate(
|
|
&Assertion::NotContains {
|
|
value: "markdown".into()
|
|
},
|
|
output
|
|
));
|
|
}
|
|
|
|
#[test]
|
|
fn exact_trims_but_does_not_accept_explanation() {
|
|
assert!(evaluate(&Assertion::Exact { value: "25".into() }, " 25\n"));
|
|
assert!(!evaluate(&Assertion::Exact { value: "25".into() }, "25 GB"));
|
|
}
|
|
|
|
#[test]
|
|
fn word_limit_counts_whitespace_tokens() {
|
|
assert!(evaluate(
|
|
&Assertion::MaxWords { value: 4 },
|
|
"one two three four"
|
|
));
|
|
assert!(!evaluate(
|
|
&Assertion::MaxWords { value: 3 },
|
|
"one two three four"
|
|
));
|
|
}
|
|
}
|