ALMA-13B
LLaMA-2-based translator that beats NLLB-54B and GPT-3.5 using monolingual pretraining plus a small set of human-written parallel data.
About
ALMA (Advanced Language Model-based trAnslator) is a research project from Haoran Xu with Microsoft collaborators that rethinks how translation LLMs are trained. Instead of depending on massive parallel corpora, it does two stages: full-weight fine-tuning on monolingual data first, then fine-tuning on a small set of high-quality human-written parallel data. The follow-up ALMA-R adds contrastive preference optimization, and X-ALMA extends the approach to 50 languages with plug-and-play language modules.
The headline result from the ICLR 2024 paper: ALMA-13B gains over 12 BLEU and 12 COMET points over its zero-shot baseline across WMT21 and WMT22 test sets, outperforming NLLB-54B and GPT-3.5-text-davinci-003. All checkpoints and datasets are MIT-licensed on Hugging Face, which makes ALMA one of the most commercially friendly open translation LLMs.
Specs
Context window
Pricing
Free quota
API access
Weights
Released
Benchmarks
Author-reported (ICLR 2024 paper): +12 BLEU and +12 COMET over zero-shot LLaMA-2 across 10 WMT21/WMT22 directions; superior to NLLB-54B and GPT-3.5-text-davinci-003
Provider-reported unless a source says otherwise. Benchmarks are marketing-adjacent by nature — treat them as directional.
Use cases
Trust score — 7/10
Scored 7/10 (verified tier, metrics-v1). Evidence: ICLR 2024 paper (arXiv:2309.11674) with WMT21/WMT22 BLEU and COMET gains confirmed via the paper and Hugging Face release notes; six checkpoints plus training datasets released MIT. Strengths: landmark data-efficiency result, permissive MIT license, active follow-ups (ALMA-R, X-ALMA to 50 languages). Docked for: (1) research artifact with no product or support channel; (2) 6-language core coverage in the original release, English-centric evaluation. License: MIT.
How the trust score is built
Full methodology →- Adoption · stars/downloads, log-scaled vs peers
- 7.0/10
- Maintenance · recency, releases, issue load
- 6.0/10
- Openness · license + docs
- 10.0/10
- Manual review · sanity read + red-flag check
- 7.0/10
- · ICLR 2024 paper (arXiv:2309.11674), Haoran Xu with Microsoft collaborators
- · Six checkpoints (7B/13B, LoRA, R variants) plus training datasets released
- · Follow-ups: ALMA-R (contrastive preference optimization), X-ALMA (50 languages)