feat(dataset-updater): Virtues FR→EN/RU/PT translation configs (#218 phase 2) by jsboige · Pull Request #231 · ArgumentumGames/Argumentum

jsboige · 2026-04-24T00:38:42Z

Summary

Completes phase 2 of #218. With the CSV i18n columns (PR #223) and Virtue entity scaffold (PR #229) now on master, this PR wires the DatasetUpdater to populate Virtues translations via OpenAI.

What's added

3 new configs in `DatasetUpdaterRootConfig.cs` (Enabled=false each)

Name	Target fields	Model
Translate Virtues to English by chunk empty-only 0-shot	`family_en`, `subfamily_en`, `subsubfamily_en`, `title_en`, `description_en`, `remark_en`, `link_en`	gpt-5.4-mini
Translate Virtues to Russian by chunk empty-only 0-shot	`*_ru`	gpt-5.4-mini
Translate Virtues to Portuguese by chunk empty-only 0-shot	`*_pt` (pt-PT)	gpt-5.4-mini

6 new prompt files in `DatasetUpdater/Resources/`

`PromptVirtuesTranslate{En,Ru,Pt}{User,Assistant}.txt` — each enforces:

Taxonomy coherence across entries (same family name translated consistently)
FR "vous" register → target-language equivalent (infinitive EN, impératif/infinitif RU, 2nd person formal pt-PT)
Canonical logic terminology
Wikipedia link localization when an equivalent page exists

Design rationale

`SelectEmptyTargets=true` — re-runs only fill gaps, so iterative passes are cheap
`ChunkSize=8`, `MaxGroupItemNb=12` — small batches because translation quality degrades on niche logic terms with large batches
Eco tier (`gpt-5.4-mini`) matching Fallacies translation configs that are already production-proven
pt-PT explicit (not pt-BR), consistent with Rules PT retranslation (Full retranslation needed for PT Rules CSV (blocked by #183) #211)

Test plan

`dotnet build` succeeds (0 errors)
88 tests pass
End-to-end: set `Enabled=true` on Virtues EN config, run pipeline with OpenAI key on a 3-5 row subset, verify output quality and column population (same acceptance gate as the feat(translations): Upgrade OpenAI translation pipeline + complete Virtues i18n #183 DatasetUpdater SDK upgrade)

Follow-ups not in this PR

CSV data population: requires someone to run the configs with `Enabled=true` and an OpenAI key (local-only, costs money — intentionally gated)
pt-PT quality post-run: could reuse the epic(publication): text & translation finalization campaign #202 micro-fix workflow for rare idiomatic corrections

🤖 Generated with Claude Code

) Completes phase 2 of #218 by wiring the DatasetUpdater to the Virtues taxonomy now that the CSV columns (_en/_ru/_pt) and entity properties are in place (PR #223 + #229 — both merged on master). Changes: - DatasetUpdaterRootConfig.cs: 3 new configs (Enabled=false each), one per target language. Same structure as the Fallacies translation configs that have been production-proven on this pipeline - 6 new prompt files adapted for Virtues taxonomic content: - PromptVirtuesTranslate{En,Ru,Pt}{User,Assistant}.txt - Each enforces taxonomy coherence across entries, FR "vous" → target-language register, canonical logic terminology, and wiki link localization Design notes: - SelectEmptyTargets=true so re-runs only fill gaps (cheap deltas) - ChunkSize=8, MaxGroupItemNb=12 — small batches for translation quality (large batches tend to drift on niche logic terms) - Model: gpt-5.4-mini (eco tier, matches Fallacies translation configs) - pt-PT explicitly requested (not pt-BR), consistent with Rules PT retranslation (#211) Test plan: - dotnet build succeeds - 88 tests pass - End-to-end: set Enabled=true on one config, run pipeline with OpenAI key, verify 3-5 rows translate cleanly (same validation gate as #183 DatasetUpdater acceptance) Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

Adds 3 DatasetUpdaterConfig entries (Enabled=false) for translating the 76 untranslated scenarios per target language (EN/RU/PT) via OpenAI GPT-5.4-mini, plus 6 Scenarii-specific prompt files (User + Assistant x 3 languages). Audit (167 total FR-complete scenarios): - EN: 85 scenarios missing ≥1 field (category, subcategory, title, smoothTalker, drawer, context, issue, suggestion_en) - RU: 76 scenarios missing _ru fields - PT: 76 scenarios missing _pt fields Config pattern mirrors #231 (Virtues): SequentialChunks, ChunkSize=8, SelectEmptyTargets=true, MaxGroupItemNb=12, UseFunctionCalling=true. Each config uses a dedicated prompt pair tailored to scenario context: - Scenarii-specific field mapping (catégorie→category, baratineur→smoothTalker, piocheur→drawer, enjeu→issue, suggestion→suggestion_en — only one with suffix) - Narrative-tone instructions (rhythm, humor, brevity) - pt-PT (not pt-BR); Russian contemporary register (not archaic) - Category/subcategory consistency across scenarios - Measured adaptation of French cultural references Prerequisites satisfied: SDK migration #183 (PR #210) + GPT-5.x models (PR #222). Scenario.cs entity already had complete i18n scaffolding (EN without _en suffix is historical design; RU/PT use suffix). Build: 0 errors, 17 warnings (pre-existing). Tests: 88 pass, 1 skip (Freeplane GUI), 0 fail. Closes #219 (pending successful end-to-end run with OpenAI key). Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

…anslations + fix BOM bug + csproj wildcards (#236) Closes #236. ## Summary ### Prompt refinement (#236) - Fix hallucinated "Vous → impératif" instruction (FR descriptions are declarative/nominal, not 2nd person) - Add canonical family mapping (9 EN / 12 PT / 12 RU families with fixed translations) - Wikipedia link caveat (no fabricated URLs, leave empty if uncertain) - Align assistant prompts with corrected user prompts ### csproj copy gap fix (#236) - Replace 21 explicit `<None Update>` entries with 2 wildcards (`*.txt`, `*.json`) - Fixes latent bug from #228/#231/#232: 21 resource files (Virtues/Scenarii/Rules prompts, CSV samples, Fallacies JSON) were not copied to bin output, blocking DatasetUpdater runs ### Virtues taxonomy translation campaign - EN: 218/223 (97.8%), 1468 _en fields - PT: 223/223 (100%), 1503 _pt fields (links cleared due to hallucination risk — see follow-up note) - RU: 220/223 (98.7%), 1480 _ru fields - All 8 family translations consistent across each language ### TokenManager fix - SharpToken throws on unknown models (e.g. gpt-5.4-mini) — added try/catch with cl100k_base fallback ### Critical bug fix: BOM accumulation in DatasetUpdater - `UtilityExtensions.cs`: strip BOM on read with `TrimStart('')` - `DataSetInfo.cs`: write without BOM via `new UTF8Encoding(encoderShouldEmitUTF8Identifier: false)` - CSV: stripped 3 accumulated BOMs from header (1 per translation cycle EN/PT/RU) - Without fix: CsvHelper would silently mismatch `pk` column header → empty PK column → broken `SelectEmptyTargets` idempotency for all future DatasetUpdater runs ### Review feedback follow-ups 1. EN subsubfamily inconsistency — pk=4,5 "Actual argument" → "Real argument" (matching pk=3, canonical mapping) 2. PT links cleared (203 URLs) — conservative cleanup vs hallucination risk, can be regenerated via future PT run with corrected prompt + caveat ## Validation - ✅ Build: 0 errors, 17 warnings (identical to master) - ✅ Tests: 88 pass / 0 fail / 1 skip (identical to master) - ✅ CSV: 0 BOM, header `pk` clean - ✅ EN consistency: 3× "Real argument", 0× "Actual argument" 🤖 Generated with [Claude Code](https://claude.com/claude-code)

clusterManager-Myia · 2026-05-03T21:54:57Z

Post-hoc review — PR #231 feat(dataset-updater): Virtues FR→EN/RU/PT translation configs (#218 phase 2)

Verdict: ⚠️ Mostly LGTM — prompt issues noted (fixed in #238)

This PR wires the DatasetUpdater to populate Virtues translations via OpenAI function calling. Three new configs (EN, RU, PT) + 6 prompt template files.

Positive:

Configs are Enabled = false — safe default.
SelectEmptyTargets = true — only fills blank fields, won't overwrite existing translations.
AutoCompare = true with title_fr/title_en comparison — smart quality gate.
ChunkSize = 8 for Virtues (smaller dataset than Fallacies) is appropriate.

Prompt concerns (noted as fixed in #238):

EN prompt incorrectly states descriptions use "vous" (2nd person) — PR description says they're actually declarative/nominal. This would cause the LLM to produce wrong register.
RU prompt also references imperative/infinitive style which is incorrect.
No canonical family mapping provided — LLM might produce inconsistent taxonomy names.
Wikipedia link instruction is too aggressive ("cherche l'équivalent") — risks hallucinated URLs.

Since #238 fixes these prompt issues, the overall trajectory is correct, but this PR in isolation has prompt quality problems that could produce bad translations if run as-is.

Verdict adjusted: Acceptable as a scaffold, but should not be run without #238's prompt fixes applied.

jsboige merged commit 108840a into master Apr 24, 2026
1 check passed

jsboige mentioned this pull request Apr 24, 2026

feat(dataset-updater): Scenarii FR→EN/RU/PT translation configs (#219) #232

Merged

5 tasks

This was referenced Apr 25, 2026

chore(dataset-updater): refine Virtues EN translation prompt + fix csproj copy gap #236

Closed

chore(dataset-updater): refine Virtues EN prompt + fix csproj copy gap (#236) #238

Merged

jsboige mentioned this pull request May 23, 2026

fix(fallacies): clean corrupted/contaminated cells in link_fr/link_pt columns #334

Closed

jsboige deleted the feat/virtues-translation-configs branch June 1, 2026 21:42

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Uh oh!

feat(dataset-updater): Virtues FR→EN/RU/PT translation configs (#218 phase 2) - #231

feat(dataset-updater): Virtues FR→EN/RU/PT translation configs (#218 phase 2)#231
jsboige merged 1 commit into
masterfrom
feat/virtues-translation-configs

jsboige commented Apr 24, 2026

Uh oh!

Uh oh!

clusterManager-Myia commented May 3, 2026

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

2 participants

Uh oh!

Conversation

jsboige commented Apr 24, 2026

Summary

What's added

3 new configs in `DatasetUpdaterRootConfig.cs` (Enabled=false each)

6 new prompt files in `DatasetUpdater/Resources/`

Design rationale

Test plan

Follow-ups not in this PR

Uh oh!

Uh oh!

clusterManager-Myia commented May 3, 2026

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

2 participants