umeiko/Haidass-Translate-143M-eval

Haidass-Translate-143M Evaluation Assets

English | 中文

Evaluation assets for DALabCommunity/Haidass-Translate-143M and 19 comparison models on FLORES-200 dev and FLORES+ devtest (Chinese–English, both directions): per-sentence predictions, score tables for both benchmarks, decontamination audits, and the scripts needed to recompute every reported number.

Evaluation setup

• Benchmarks: FLORES-200 dev (997 sentences per direction) and FLORES+ devtest (1,012 sentences per direction, zero overlap with dev); language pair eng_Latn ↔ zho_Hans
  • Maintained release: openlanguagedata/flores_plus (FLORES+, maintained by OLDI; gated, access is auto-approved after accepting the terms; Simplified Chinese is now coded cmn_Hans)
  • Original archive: facebook/flores (unmaintained); login-free mirror: facebookresearch/flores
• Decoding: greedy (do_sample=False), max_new_tokens=512, bf16
• Metrics: sacreBLEU corpus BLEU (tokenize=zh for Chinese targets, tokenize=13a for English targets) and chrF++ (word_order=2)
• Prompt rule: each model is evaluated with prompts byte-identical to its training or official template — chat template for LLMs, forced-BOS for NLLB / M2M / OPUS-MT

Repository structure

Path	Contents
ZEROSHOT_RESULTS.md	Score tables for both benchmarks (20 models), metric definitions, reading notes, ablation analysis and sample spot checks
audit_report.json	n-gram decontamination audit, 8M training data vs FLORES-200 dev (top-50 overlapping samples included)
audit_floresplus_devtest.json	Same audit vs FLORES+ devtest
audit_floresplus_devtest_stepfun.json	Same audit restricted to the 157,593 general-domain (STEP_FUN) samples used in the mix variant, vs FLORES+ devtest
predictions/	Per-sentence predictions: one pred_<model>_flores_dev.jsonl / pred_<model>_floresplus_devtest.jsonl per model per benchmark
scripts/eval_translate.py	LLM evaluation script (chat template; --eval-dir selects the benchmark)
scripts/eval_translate_seq2seq.py	Seq2seq evaluation script (NLLB / M2M / OPUS-MT)
scripts/compute_metrics.py	Metric recomputation (BLEU / chrF++)
scripts/ngram_audit.py	Decontamination audit script

Prediction file format

One JSON object per line:

{"id": 0, "direction": "en-zh", "src": "...", "ref": "...", "hyp": "..."}

direction is en-zh or zh-en; src is the source sentence, ref the reference translation, hyp the model output.

Recomputing the metrics

pip install sacrebleu
python scripts/compute_metrics.py predictions/pred_haidass-8M-trans_floresplus_devtest.jsonl
# == predictions/pred_haidass-8M-trans_floresplus_devtest.jsonl
#   en-zh: n=1012 BLEU=25.06 chrF++=17.72
#   zh-en: n=1012 BLEU=13.68 chrF++=40.02

Summary of findings

• All 20 model rows were re-evaluated on FLORES+ devtest, which has zero overlap with dev. Both tables and all per-sentence predictions are published here for inspection. The released model's scores move within +1.2 / -1.0 BLEU between the two benchmarks (dev 23.89 / 14.65 to devtest 25.06 / 13.68).
• Drafter-8M control (same data and recipe at 8M scale, but trained from random initialization): en→zh 12.04 / zh→en 5.47, versus 23.89 / 14.65 when starting from the pretrained base. Without the base, scaling the same parallel SFT data from 1M to 8M adds only +1.2 / +0.3 BLEU — base-model pretraining contributes language knowledge that additional SFT parallel data does not replace.
• With the pretrained base, en→zh BLEU rises monotonically with data scale: 16.65 → 20.10 → 23.64 → 23.89 (1M / 4M / 8M-mix / 8M translation-only), while zh→en plateaus around 14.4. The 8M-mix-v1 collapse (zh→en 7.34) was traced to uncleaned general-domain data and missing inter-document attention isolation during packing; both were fixed in v2. Full analysis: ZEROSHOT_RESULTS.md.

Decontamination

To check whether the reported scores are inflated by test-set leakage, every training sample was cut into consecutive fragments (8 words for English, 10 characters for Chinese), and any sample sharing at least one fragment with any test sentence was counted as a hit. Of the 15,830,983 training samples, 1,147 (0.0072%) hit FLORES-200 dev and 1,788 (0.0113%) hit FLORES+ devtest; the STEP_FUN-only subset shows 6 hits (0.0038%) against FLORES+ devtest. Manual inspection of the most-overlapping cases shows common-phrase-level overlap rather than full-sentence leakage. Each audit report includes its 50 most-overlapping samples for independent review.