Modelpublic

openeurollm/OLMo-3-7B-Dolci-Translated-A-25EN

sourceHugging Faceapache-2.0updated 4mo agoView on Hugging Face

0likes29downloads

Model Card

OLMo-3-7B Dolci-Translated A-25EN (matched compute)

Continued-SFT of openeurollm/OLMo-3-7B-Instruct-SFT on a $25/75$ English:EU mixture, the EU-leaning configuration of the paper Translate, Replay, Mix: Exploring Multilingual Post-Training for Low-Resource European Languages. This is the matched-compute variant: both A-75EN and A-25EN consume the same $2.87$M samples × $2$ epochs at the same $1$M-token batch size; A-25EN takes 5398 steps to A-75EN's 3998 because the gemma-translated EU text packs more tokens per sample.

—Qualitative completions viewer (this checkpoint's progression is on the right): https://ferreirafabio.github.io/olmo3-multilingual-dolci-sft-progression/

Recipe


Base checkpoint	`openeurollm/OLMo-3-7B-Instruct-SFT` (our reproduction at parity with `allenai/OLMo-3-7B-Instruct-SFT`)
English half (Dolci replay)	`allenai/Dolci-Instruct-SFT`, 25% of the mixture
EU half (Dolci-Translated)	`openeurollm/Dolci-Instruct-SFT-translated`, 75% of the mixture, 7 EU languages translated with `gemma-3-27b-it`
EU languages	cs, de, es, fi, fr, it, sv
Total samples	4.62M (1,155,000 en + 3,465,000 EU)
Final step	5398 (matched-compute variant)
Chat template	`olmo` (inherited from base)

Training configuration

Identical to openeurollm/OLMo-3-7B-Dolci-Translated-A-75EN: AdamW, peak LR $8\times10^{-5}$, batch ${\sim}1$M tokens, seq_len 32k, BF16, DeepSpeed ZeRO 2, $8\times$H200 SXM. Only the English:EU ratio differs.

Evaluation

Bradley-Terry Elo (Qwen3.5-27B judge, 500 battles/language, 100 bootstraps):

Metric	A-25EN (this checkpoint)	Baseline (`openeurollm/OLMo-3-7B-Instruct-SFT`)
Overall Elo	$782 \pm 6$	$762 \pm 7$
English Elo	$913 \pm 16$	$954 \pm 16$
Non-English Elo	$\mathbf{742 \pm 8}$	$697 \pm 9$

Per-language Elo (cs / de / es / fi / fr / it / sv):

en	cs	de	es	fi	fr	it	sv
$913 \pm 16$	$\mathbf{745 \pm 15}$	$\mathbf{701 \pm 23}$	$\mathbf{763 \pm 18}$	$\mathbf{813 \pm 33}$	$\mathbf{756 \pm 17}$	$801 \pm 15$	$756 \pm 33$

A-25EN dominates the EU columns (best on cs, de, es, fi, fr) at the cost of $41$ English Elo points. Pick this checkpoint when the deployment is EU-facing and a moderate English regression is acceptable.

Intermediate checkpoints

Training-step revisions (step500, step1500, step2500, step3500) are available as HF git revisions of this repo (loaded via revision="step1500") and back the qualitative completions viewer at https://ferreirafabio.github.io/olmo3-multilingual-dolci-sft-progression/.

How to load

python

from transformers import AutoTokenizer, AutoModelForCausalLM
tok = AutoTokenizer.from_pretrained("openeurollm/OLMo-3-7B-Dolci-Translated-A-25EN")
model = AutoModelForCausalLM.from_pretrained("openeurollm/OLMo-3-7B-Dolci-Translated-A-25EN", torch_dtype="bfloat16")

Citation

Please cite the paper and the OLMo-3 family if you use this checkpoint.