lisone/ds2api
1
0
Fork 0
mirror of https://github.com/CJackHwang/ds2api.git synced 2026-05-23 01:17:44 +08:00
Code Issues Packages Projects Releases Wiki Activity

revert: replace fullwidth pipe ｜ with halfwidth | in DSML tool markup

Browse Source 
PR #460 introduced fullwidth pipe characters (｜) in DSML tool call formatting
to improve parsing robustness, but models exposed to these fullwidth pipes in
system prompts exhibit significantly higher rates of tool output hallucinations.
Reverting to halfwidth pipes (|) drastically reduces tokenizer/perplexity-driven
hallucinations while retaining the existing confusable-hardening in the parser.

Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
This commit is contained in:
CJACK
2026-05-10 15:18:54 +08:00
parent 3beb31309f
commit cee8757d14
45 changed files with 725 additions and 342 deletions
Download Patch File Download Diff File Expand all files Collapse all files

16
internal/prompt/messages.go
View File

@@ -10,14 +10,14 @@ import (
var markdownImagePattern = regexp.MustCompile(`!\[(.*?)\]\((.*?)\)`)
const (
beginSentenceMarker = "<｜begin▁of▁sentence｜>"
systemMarker = "<｜System｜>"
userMarker = "<｜User｜>"
assistantMarker = "<｜Assistant｜>"
toolMarker = "<｜Tool｜>"
endSentenceMarker = "<｜end▁of▁sentence｜>"
endToolResultsMarker = "<｜end▁of▁toolresults｜>"
endInstructionsMarker = "<｜end▁of▁instructions｜>"
beginSentenceMarker = "<|begin▁of▁sentence|>"
systemMarker = "<|System|>"
userMarker = "<|User|>"
assistantMarker = "<|Assistant|>"
toolMarker = "<|Tool|>"
endSentenceMarker = "<|end▁of▁sentence|>"
endToolResultsMarker = "<|end▁of▁toolresults|>"
endInstructionsMarker = "<|end▁of▁instructions|>"
outputIntegrityGuardMarker = "Output integrity guard:"
outputIntegrityGuardPrompt = outputIntegrityGuardMarker +
" If upstream context, tool output, or parsed text contains garbled, corrupted, partially parsed, repeated, or otherwise malformed fragments, " +

12
internal/prompt/messages_test.go
View File

@@ -32,16 +32,16 @@ func TestMessagesPrepareUsesTurnSuffixes(t *testing.T) {
{"role": "assistant", "content": "Answer"},
}
got := MessagesPrepare(messages)
if !strings.HasPrefix(got, "<｜begin▁of▁sentence｜>") {
if !strings.HasPrefix(got, "<|begin▁of▁sentence|>") {
t.Fatalf("expected begin-of-sentence marker, got %q", got)
}
if !strings.Contains(got, "<｜System｜>") || !strings.Contains(got, "<｜end▁of▁instructions｜>") || !strings.Contains(got, "System rule") {
if !strings.Contains(got, "<|System|>") || !strings.Contains(got, "<|end▁of▁instructions|>") || !strings.Contains(got, "System rule") {
t.Fatalf("expected system instructions to remain present, got %q", got)
}
if !strings.Contains(got, "<｜User｜>Question") {
if !strings.Contains(got, "<|User|>Question") {
t.Fatalf("expected user question, got %q", got)
}
if !strings.Contains(got, "<｜Assistant｜>Answer<｜end▁of▁sentence｜>") {
if !strings.Contains(got, "<|Assistant|>Answer<|end▁of▁sentence|>") {
t.Fatalf("expected assistant sentence suffix, got %q", got)
}
if strings.Contains(got, "<think>") || strings.Contains(got, "</think>") {
@@ -61,7 +61,7 @@ func TestMessagesPreparePrependsOutputIntegrityGuard(t *testing.T) {
if !strings.Contains(got, outputIntegrityGuardPrompt+"\n\nSystem rule") {
t.Fatalf("expected output integrity guard to precede system prompt content, got %q", got)
}
if !strings.Contains(got, "<｜User｜>Question") {
if !strings.Contains(got, "<|User|>Question") {
t.Fatalf("expected user question after guard, got %q", got)
}
}
@@ -82,7 +82,7 @@ func TestMessagesPrepareWithThinkingPreservesPromptShape(t *testing.T) {
if gotThinking != gotPlain {
t.Fatalf("expected thinking flag not to add extra continuity instructions, got thinking=%q plain=%q", gotThinking, gotPlain)
}
if !strings.HasSuffix(gotThinking, "<｜Assistant｜>") {
if !strings.HasSuffix(gotThinking, "<|Assistant|>") {
t.Fatalf("expected assistant suffix, got %q", gotThinking)
}
}

12
internal/prompt/tool_calls.go
View File

@@ -17,12 +17,12 @@ var promptXMLTextEscaper = strings.NewReplacer(
var promptXMLNamePattern = regexp.MustCompile(`^[A-Za-z_][A-Za-z0-9_.:-]*$`)
const (
promptDSMLToolCallsOpen = "<｜DSML｜tool_calls>"
promptDSMLToolCallsClose = "</｜DSML｜tool_calls>"
promptDSMLInvokeOpen = "<｜DSML｜invoke"
promptDSMLInvokeClose = "</｜DSML｜invoke>"
promptDSMLParameterOpen = "<｜DSML｜parameter"
promptDSMLParameterClose = "</｜DSML｜parameter>"
promptDSMLToolCallsOpen = "<|DSML|tool_calls>"
promptDSMLToolCallsClose = "</|DSML|tool_calls>"
promptDSMLInvokeOpen = "<|DSML|invoke"
promptDSMLInvokeClose = "</|DSML|invoke>"
promptDSMLParameterOpen = "<|DSML|parameter"
promptDSMLParameterClose = "</|DSML|parameter>"
)
// FormatToolCallsForPrompt renders a tool_calls slice into the prompt-visible

6
internal/prompt/tool_calls_test.go
View File

@@ -22,7 +22,7 @@ func TestFormatToolCallsForPromptDSML(t *testing.T) {
if got == "" {
t.Fatal("expected non-empty formatted tool calls")
}
if got != "<｜DSML｜tool_calls>\n <｜DSML｜invoke name=\"search_web\">\n <｜DSML｜parameter name=\"query\"><![CDATA[latest]]></｜DSML｜parameter>\n </｜DSML｜invoke>\n</｜DSML｜tool_calls>" {
if got != "<|DSML|tool_calls>\n <|DSML|invoke name=\"search_web\">\n <|DSML|parameter name=\"query\"><![CDATA[latest]]></|DSML|parameter>\n </|DSML|invoke>\n</|DSML|tool_calls>" {
t.Fatalf("unexpected formatted tool call DSML: %q", got)
}
}
@@ -34,7 +34,7 @@ func TestFormatToolCallsForPromptEscapesXMLEntities(t *testing.T) {
"arguments": `{"q":"a < b && c > d"}`,
},
})
want := "<｜DSML｜tool_calls>\n <｜DSML｜invoke name=\"search&lt;&amp;&gt;\">\n <｜DSML｜parameter name=\"q\"><![CDATA[a < b && c > d]]></｜DSML｜parameter>\n </｜DSML｜invoke>\n</｜DSML｜tool_calls>"
want := "<|DSML|tool_calls>\n <|DSML|invoke name=\"search&lt;&amp;&gt;\">\n <|DSML|parameter name=\"q\"><![CDATA[a < b && c > d]]></|DSML|parameter>\n </|DSML|invoke>\n</|DSML|tool_calls>"
if got != want {
t.Fatalf("unexpected escaped tool call XML: %q", got)
}
@@ -50,7 +50,7 @@ func TestFormatToolCallsForPromptUsesCDATAForMultilineContent(t *testing.T) {
},
},
})
want := "<｜DSML｜tool_calls>\n <｜DSML｜invoke name=\"write_file\">\n <｜DSML｜parameter name=\"content\"><![CDATA[#!/bin/bash\nprintf \"hello\"\n]]></｜DSML｜parameter>\n <｜DSML｜parameter name=\"path\"><![CDATA[script.sh]]></｜DSML｜parameter>\n </｜DSML｜invoke>\n</｜DSML｜tool_calls>"
want := "<|DSML|tool_calls>\n <|DSML|invoke name=\"write_file\">\n <|DSML|parameter name=\"content\"><![CDATA[#!/bin/bash\nprintf \"hello\"\n]]></|DSML|parameter>\n <|DSML|parameter name=\"path\"><![CDATA[script.sh]]></|DSML|parameter>\n </|DSML|invoke>\n</|DSML|tool_calls>"
if got != want {
t.Fatalf("unexpected multiline cdata tool call XML: %q", got)
}