Files
picoclaw/cmd/membench/eval.go
T
BeaconCat f1b659e5ef membench: add LLM-as-Judge evaluation mode (#2484)
* membench: add LLM-as-Judge evaluation mode

Add --eval-mode=llm to membench for LLM-based answer generation and
semantic scoring via an OpenAI-compatible API endpoint.

New files:
- llm_client.go: generic OpenAI-compatible chat completion client
  with support for API key, configurable timeout, and optional
  chat_template_kwargs (for llama.cpp thinking models)
- eval_llm.go: LLM answer generation + LLM-as-Judge scoring for
  both legacy and seahorse retrieval modes

Changes to main.go:
- --eval-mode flag (token|llm) to select evaluation strategy
- --api-base, --api-key, --model flags with env var fallback
  (MEMBENCH_API_BASE, MEMBENCH_API_KEY, MEMBENCH_MODEL)
- --no-thinking flag for llama.cpp + Qwen thinking models
- --limit flag to cap QA questions per sample for quick testing

* style: fix golangci-lint formatting (gofmt + golines)

* fix: address Copilot review feedback

- Validate --model is required for LLM eval mode
- Use rune-based truncation to preserve valid UTF-8
- Precompute totalQA count outside inner loop
- Log SearchMessages errors instead of silently skipping

* fix: address Copilot review round 2

- Validate --eval-mode accepts only 'token' or 'llm'
- Normalize base URL to avoid /v1/v1 duplication
- Separate token/LLM results for correct PrintComparison labeling
- Log ExpandMessages errors instead of silently ignoring
- Short-circuit with 0 scores when no context retrieved (match token eval)
- Add --timeout flag wired to LLMClientOptions.Timeout

* fix: address review P1+P2 — sort alignment, failure sentinel, score parser

- P1: Replace hand-rolled sortByRank with sort.Slice (ascending, best
  first) matching eval.go's EvalSeahorse — ensures BudgetTruncate keeps
  best-ranked messages when truncation occurs
- P2: Use -1.0 sentinel for LLM API failures and parse errors, distinct
  from genuine 0.0 score; aggregateMetrics skips -1.0 entries for F1
  averaging while still counting HitRate
- P2: Use regexp \b([1-5])\b for judge score extraction instead of
  first-digit scan — avoids misparses on '5/5', 'Score: 3' etc.

* fix: address Copilot review round 2

- Fix F1/HitRate weighted aggregation: track ValidF1Count separately so
  computeModeAgg weights F1 by valid scores only, not TotalQuestions
- No-context retrieval failure uses 0.0 (genuine bad score) instead of
  -1.0 sentinel (reserved for API/parse failures)
- Validate --timeout > 0 to prevent disabling HTTP timeouts

* fix: remove hardcoded /v1 from API base URL

Users now provide the full versioned path in --api-base (e.g. /v1, /v4).
Code only appends /chat/completions. Default changed to
http://127.0.0.1:8080/v1 for backward compatibility.

* fix: address Copilot review round 3

- ValidF1Count=0 when all scores are sentinel (no forced =1)
- Backward compat: old eval JSON without ValidF1Count falls back to
  TotalQuestions in computeModeAgg
- Skip empty section in PrintComparison when tokenResults is empty
- Update --api-base flag help to document /v1 default and version path
- Add sentinel aggregation unit tests (partial, all, weighted)

* feat: add --retries flag with exponential backoff for transient LLM errors

Retry on timeout, 5xx, and 429 (rate limit) with 1s/2s/4s backoff.
Default 3 retries, configurable via --retries. Context cancellation
is respected between retries.

* fix: address Copilot review round 4

- runReport splits results by mode suffix into token/llm for PrintComparison
- backward compat fallback (ValidF1Count=0 -> TotalQuestions) only for
  non-LLM modes; LLM modes keep ValidF1Count=0 when all scores sentinel
- MaxRetries==0 means no retry; only negative falls back to default 3
- truncateStr uses []rune to avoid cutting multi-byte UTF-8 characters
- Complete() returns error on empty LLM response (vs silent empty string)

* feat: --no-thinking adapts to llama.cpp, Ollama, and GLM backends

Send all three disable-thinking fields simultaneously:
- chat_template_kwargs.enable_thinking=false (llama.cpp, GLM)
- think=false (Ollama 0.9+)
- thinking.type=disabled (GLM/Zhipu)
Each backend picks the field it recognizes and ignores the rest.
Also bumps max_tokens from 512 to 2048 for thinking models.

* feat: mixed model eval + concurrent QA workers

- Add --judge-model, --judge-api-base, --judge-api-key flags for separate judge model
- Add --concurrency flag (default 1) with semaphore-based goroutine pool
- Add reasoning_content fallback for GLM/DeepSeek style responses
- Prepend /no_think to system prompt for Ollama /v1 compatibility
- Reduce default MaxTokens from 2048 to 512 (answers are 1-3 sentences)
- Extract evalQAWorker and buildSeahorseContext for shared concurrent logic

---------

Co-authored-by: BeaconCat <BeaconCat@users.noreply.github.com>
2026-04-15 21:15:17 +08:00

413 lines
11 KiB
Go

package main
import (
"context"
"encoding/json"
"fmt"
"log"
"os"
"path/filepath"
"sort"
"strings"
"github.com/sipeed/picoclaw/pkg/seahorse"
)
// EvalResult holds per-sample evaluation results for one mode.
type EvalResult struct {
Mode string `json:"mode"`
SampleID string `json:"sampleId"`
QAResults []QAResult `json:"qaResults"`
Agg AggMetrics `json:"aggregated"`
}
// QAResult holds metrics for a single QA pair.
type QAResult struct {
Question string `json:"question"`
Category int `json:"category"`
GoldAnswer string `json:"goldAnswer"`
TokenF1 float64 `json:"tokenF1"`
HitRate float64 `json:"hitRate"`
}
// AggMetrics holds aggregated evaluation metrics.
type AggMetrics struct {
OverallF1 float64 `json:"overallF1"`
OverallHitRate float64 `json:"overallHitRate"`
ByCategory map[int]*CatMetrics `json:"byCategory"`
TotalQuestions int `json:"totalQuestions"`
ValidF1Count int `json:"validF1Count"`
}
// CatMetrics holds metrics for a single category.
type CatMetrics struct {
F1 float64 `json:"f1"`
HitRate float64 `json:"hitRate"`
QuestionCount int `json:"questionCount"`
ValidF1Count int `json:"validF1Count"`
}
// EvalLegacy evaluates using legacy session store (raw history + budget truncation).
func EvalLegacy(
ctx context.Context,
samples []LocomoSample,
legacy *LegacyStore,
budgetTokens int,
) []EvalResult {
results := make([]EvalResult, 0, len(samples))
for si := range samples {
sample := &samples[si]
history := legacy.GetHistory(sample.SampleID)
// Convert messages to content strings
allContent := make([]string, 0, len(history))
for _, msg := range history {
allContent = append(allContent, msg.Content)
}
qaResults := make([]QAResult, 0, len(sample.QA))
for qi := range sample.QA {
qa := &sample.QA[qi]
// Budget truncate the full history
truncated, _ := BudgetTruncate(allContent, budgetTokens)
context := StringListToContent(truncated)
f1 := TokenOverlapF1(context, qa.AnswerString())
hitRate := RecallHitRate(qa.Evidence, sample, context)
qaResults = append(qaResults, QAResult{
Question: qa.Question,
Category: qa.Category,
GoldAnswer: qa.AnswerString(),
TokenF1: f1,
HitRate: hitRate,
})
}
results = append(results, EvalResult{
Mode: "legacy",
SampleID: sample.SampleID,
QAResults: qaResults,
Agg: aggregateMetrics(qaResults),
})
}
return results
}
// EvalSeahorse evaluates using seahorse short memory (per-keyword search + expand).
func EvalSeahorse(
ctx context.Context,
samples []LocomoSample,
ir *SeahorseIngestResult,
budgetTokens int,
) []EvalResult {
store := ir.Engine.GetRetrieval().Store()
retrieval := ir.Engine.GetRetrieval()
results := make([]EvalResult, 0, len(samples))
for si := range samples {
sample := &samples[si]
convID, ok := ir.ConvMap[sample.SampleID]
if !ok {
log.Printf("WARN: no conversation ID for sample %s", sample.SampleID)
continue
}
qaResults := make([]QAResult, 0, len(sample.QA))
for qi := range sample.QA {
qa := &sample.QA[qi]
keywords := ExtractKeywords(qa.Question)
// Search each keyword individually and union results,
// tracking best BM25 rank per message for relevance sorting.
bestRank := map[int64]float64{}
for _, kw := range keywords {
searchResults, err := store.SearchMessages(ctx, seahorse.SearchInput{
Pattern: kw,
ConversationID: convID,
Limit: 20,
})
if err != nil {
log.Printf("WARN: search failed for keyword %q: %v", kw, err)
continue
}
for _, sr := range searchResults {
if sr.MessageID > 0 {
if prev, ok := bestRank[sr.MessageID]; !ok || sr.Rank < prev {
bestRank[sr.MessageID] = sr.Rank
}
}
}
}
// Sort messageIDs by rank ascending (best/most-negative first).
// BudgetTruncate walks from the front, keeping best-ranked messages.
// Note: SQLite FTS5 bm25() returns negative values where more
// negative = better match.
messageIDs := make([]int64, 0, len(bestRank))
for id := range bestRank {
messageIDs = append(messageIDs, id)
}
sort.Slice(messageIDs, func(i, j int) bool {
return bestRank[messageIDs[i]] < bestRank[messageIDs[j]]
})
// Expand messages to get full content
var contentParts []string
if len(messageIDs) > 0 {
expandResult, err := retrieval.ExpandMessages(ctx, messageIDs)
if err != nil {
log.Printf("WARN: expand failed for sample %s: %v", sample.SampleID, err)
} else {
for _, msg := range expandResult.Messages {
contentParts = append(contentParts, msg.Content)
}
}
}
if len(contentParts) == 0 {
qaResults = append(qaResults, QAResult{
Question: qa.Question,
Category: qa.Category,
GoldAnswer: qa.AnswerString(),
TokenF1: 0.0,
HitRate: 0.0,
})
continue
}
// Budget truncate (drop worst-ranked)
truncated, _ := BudgetTruncate(contentParts, budgetTokens)
context := StringListToContent(truncated)
f1 := TokenOverlapF1(context, qa.AnswerString())
hitRate := RecallHitRate(qa.Evidence, sample, context)
qaResults = append(qaResults, QAResult{
Question: qa.Question,
Category: qa.Category,
GoldAnswer: qa.AnswerString(),
TokenF1: f1,
HitRate: hitRate,
})
}
results = append(results, EvalResult{
Mode: "seahorse",
SampleID: sample.SampleID,
QAResults: qaResults,
Agg: aggregateMetrics(qaResults),
})
}
return results
}
// aggregateMetrics computes overall and per-category metrics.
func aggregateMetrics(qaResults []QAResult) AggMetrics {
type catAccum struct {
f1Sum float64
f1Count int
hitRateSum float64
hitRateCount int
}
byCatAcc := map[int]*catAccum{}
totalF1 := 0.0
totalHitRate := 0.0
validF1Count := 0
for _, qr := range qaResults {
// Skip sentinel -1.0 scores (LLM API/parse failures) from F1 averaging.
if qr.TokenF1 >= 0 {
totalF1 += qr.TokenF1
validF1Count++
}
totalHitRate += qr.HitRate
acc, ok := byCatAcc[qr.Category]
if !ok {
acc = &catAccum{}
byCatAcc[qr.Category] = acc
}
if qr.TokenF1 >= 0 {
acc.f1Sum += qr.TokenF1
acc.f1Count++
}
acc.hitRateSum += qr.HitRate
acc.hitRateCount++
}
nHit := len(qaResults)
if nHit == 0 {
nHit = 1
}
byCat := map[int]*CatMetrics{}
for cat, acc := range byCatAcc {
cm := &CatMetrics{
QuestionCount: acc.hitRateCount,
ValidF1Count: acc.f1Count,
}
if acc.f1Count > 0 {
cm.F1 = acc.f1Sum / float64(acc.f1Count)
}
if acc.hitRateCount > 0 {
cm.HitRate = acc.hitRateSum / float64(acc.hitRateCount)
}
byCat[cat] = cm
}
var overallF1 float64
if validF1Count > 0 {
overallF1 = totalF1 / float64(validF1Count)
}
return AggMetrics{
OverallF1: overallF1,
OverallHitRate: totalHitRate / float64(nHit),
ByCategory: byCat,
TotalQuestions: len(qaResults),
ValidF1Count: validF1Count,
}
}
// SaveResults writes per-sample eval results to JSON files.
func SaveResults(results []EvalResult, outDir string) error {
if err := os.MkdirAll(outDir, 0o755); err != nil {
return fmt.Errorf("create output dir: %w", err)
}
for _, r := range results {
path := filepath.Join(outDir, fmt.Sprintf("eval_%s_%s.json", r.Mode, r.SampleID))
data, err := json.MarshalIndent(r, "", " ")
if err != nil {
return fmt.Errorf("marshal result: %w", err)
}
if err := os.WriteFile(path, data, 0o644); err != nil {
return fmt.Errorf("write result: %w", err)
}
}
return nil
}
// SaveAggregated writes a combined results.json with all modes.
func SaveAggregated(results []EvalResult, outDir string) error {
byMode := map[string][]EvalResult{}
for _, r := range results {
byMode[r.Mode] = append(byMode[r.Mode], r)
}
aggMap := map[string]AggMetrics{}
for mode, modeResults := range byMode {
aggMap[mode] = computeModeAgg(modeResults)
}
data, err := json.MarshalIndent(aggMap, "", " ")
if err != nil {
return err
}
return os.WriteFile(filepath.Join(outDir, "results.json"), data, 0o644)
}
// computeModeAgg aggregates results for a single mode using weighted averaging
// (weighted by question count per sample). All modes must have the same Mode field.
func computeModeAgg(results []EvalResult) AggMetrics {
agg := AggMetrics{ByCategory: map[int]*CatMetrics{}}
for _, r := range results {
// Backward compat: old eval JSON (token mode) without ValidF1Count → use TotalQuestions.
// LLM modes may legitimately have ValidF1Count==0 (all failures).
vf1 := r.Agg.ValidF1Count
if vf1 == 0 && r.Agg.TotalQuestions > 0 && !strings.HasSuffix(r.Mode, "-llm") {
vf1 = r.Agg.TotalQuestions
}
agg.OverallF1 += r.Agg.OverallF1 * float64(vf1)
agg.OverallHitRate += r.Agg.OverallHitRate * float64(r.Agg.TotalQuestions)
agg.TotalQuestions += r.Agg.TotalQuestions
agg.ValidF1Count += vf1
for cat, cm := range r.Agg.ByCategory {
existing, ok := agg.ByCategory[cat]
if !ok {
existing = &CatMetrics{}
agg.ByCategory[cat] = existing
}
cvf1 := cm.ValidF1Count
if cvf1 == 0 && cm.QuestionCount > 0 && !strings.HasSuffix(r.Mode, "-llm") {
cvf1 = cm.QuestionCount
}
existing.F1 += cm.F1 * float64(cvf1)
existing.HitRate += cm.HitRate * float64(cm.QuestionCount)
existing.QuestionCount += cm.QuestionCount
existing.ValidF1Count += cvf1
}
}
if agg.ValidF1Count > 0 {
agg.OverallF1 /= float64(agg.ValidF1Count)
}
if agg.TotalQuestions > 0 {
agg.OverallHitRate /= float64(agg.TotalQuestions)
}
for _, cat := range agg.ByCategory {
if cat.ValidF1Count > 0 {
cat.F1 /= float64(cat.ValidF1Count)
}
if cat.QuestionCount > 0 {
cat.HitRate /= float64(cat.QuestionCount)
}
}
return agg
}
// printSection prints a single comparison table section.
func printSection(title string, results []EvalResult) {
fmt.Printf("\n--- %s ---\n", title)
byMode := map[string][]EvalResult{}
for _, r := range results {
byMode[r.Mode] = append(byMode[r.Mode], r)
}
modes := map[string]AggMetrics{}
for mode, modeResults := range byMode {
modes[mode] = computeModeAgg(modeResults)
}
modeKeys := make([]string, 0, len(modes))
for k := range modes {
modeKeys = append(modeKeys, k)
}
sort.Strings(modeKeys)
// Collect all category keys across modes
catSet := map[int]bool{}
for _, agg := range modes {
for cat := range agg.ByCategory {
catSet[cat] = true
}
}
cats := make([]int, 0, len(catSet))
for cat := range catSet {
cats = append(cats, cat)
}
sort.Ints(cats)
fmt.Printf("%-10s %-8s %-8s", "Mode", "HitRate", "F1")
for _, cat := range cats {
fmt.Printf(" %-7s", fmt.Sprintf("C%d", cat))
}
fmt.Println()
fmt.Println(strings.Repeat("-", 10+8+8+7*len(cats)+8))
for _, mode := range modeKeys {
agg := modes[mode]
fmt.Printf("%-10s %-8.4f %-8.4f", mode, agg.OverallHitRate, agg.OverallF1)
for _, cat := range cats {
if cm, ok := agg.ByCategory[cat]; ok {
fmt.Printf(" %-7.4f", cm.HitRate)
} else {
fmt.Printf(" %-7s", "N/A")
}
}
fmt.Println()
}
}
// PrintComparison outputs a human-readable comparison table to stdout.
func PrintComparison(results []EvalResult, llmResults []EvalResult) {
if len(results) > 0 {
printSection("No LLM generation", results)
}
if len(llmResults) > 0 {
printSection("With LLM", llmResults)
}
}