This model is the direct preference optimization stage of the LFM2.5 TitleGen experiment. Model repository: ManhHoDinh/lfm25-titlegen-dpo. Benchmark publication timestamp: 2026-08-16T00:00:00+07:00.

Preliminary benchmark

The current model passed 293 of 300 evaluated English and Vietnamese examples (97.7%).

Model	Passed	Overall	English	Vietnamese	Evidence
DPO	293 / 300	97.7%	99.5%	94.1%	Preliminary automated EN/VI
SFT v2	291 / 300	97.0%	99.5%	92.2%	Preliminary automated EN/VI
Curriculum	289 / 300	96.3%	98.5%	92.2%	Preliminary automated EN/VI

Language coverage

Language	Samples	Passed	Rate	Status
German	0	0	0.0%	NOT_EVALUATED
English	198	197	99.5%	EVALUATED
Spanish	0	0	0.0%	NOT_EVALUATED
Filipino	0	0	0.0%	NOT_EVALUATED
French	0	0	0.0%	NOT_EVALUATED
Indonesian	0	0	0.0%	NOT_EVALUATED
Japanese	0	0	0.0%	NOT_EVALUATED
Korean	0	0	0.0%	NOT_EVALUATED
Lao	0	0	0.0%	NOT_EVALUATED
Malay	0	0	0.0%	NOT_EVALUATED
Burmese	0	0	0.0%	NOT_EVALUATED
Portuguese	0	0	0.0%	NOT_EVALUATED
Russian	0	0	0.0%	NOT_EVALUATED
Tamil	0	0	0.0%	NOT_EVALUATED
Thai	0	0	0.0%	NOT_EVALUATED
Vietnamese	102	96	94.1%	EVALUATED
Chinese	0	0	0.0%	NOT_EVALUATED

0 means no evaluated examples when the status is NOT_EVALUATED; it is not a measured zero score.

Methodology

This preliminary benchmark evaluates aggregate English and Vietnamese results with deterministic decoding (do_sample: false, max_new_tokens: 32). Automated rubric identifiers: 3-8_tu, khong_cham_cuoi, mot_dong, dung_ngon_ngu, khong_chep. The remaining contracted languages are shown explicitly as not evaluated.

Limitations

Only English and Vietnamese have evaluated examples.
Language correctness uses an automated heuristic.
No native review or blind preference evidence is included.
These aggregate automated results do not establish production readiness, causal improvement, or statistical significance.

Machine-readable results

See benchmark-report.json for the validated aggregate report.