License: Free for commercial use, with required attribution to the author's work. If you use this model, or text it generated, in a product, service or publication, credit "Translation models by Aletheia AI" and cite the accompanying paper. This model is a derivative of Google Gemma 3, so the Gemma Terms of Use and Prohibited Use Policy also apply.

gemma-3-1b-translate-rl

What this is. A research archive of earlier checkpoints of the Gemma-3-1B English to African-language translator (125 language tags, English to African language only). Each branch holds one set of weights; main holds only this README. For normal use, take the final model: BeardedMonster/gemma-3-1b-translate-best. Its card has the prompt format, code and the full per-language results.

Headline of the final model (FLORES+ devtest, 54 languages, greedy decoding): chrF++ 35.28, AfriCOMET 0.681 (40 languages).

Branches

branch what it is status
latest-m1 Later checkpoint of the same RL run as the final model; not separately evaluated earlier RL checkpoint, superseded by the final model
dpo-trial Two models in folders modelA/ and modelB/ experimental preference-tuning trial, not recommended
main README only no weights

The final model itself is published as BeardedMonster/gemma-3-1b-translate-best; it is not duplicated as a branch here.

How to load a branch

The prompt is one user turn (two spaces after the colon, the language name, never the ISO code); the model answers with a single language tag, then the translation.

import re, torch
from transformers import AutoTokenizer, AutoModelForCausalLM

repo = "BeardedMonster/gemma-3-1b-translate-rl"
tok = AutoTokenizer.from_pretrained("BeardedMonster/gemma-3-1b-translate-best")
model = AutoModelForCausalLM.from_pretrained(repo, revision="latest-m1", dtype=torch.bfloat16).cuda().eval()
# for the trial models: revision="dpo-trial", subfolder="modelA"  (or "modelB")

msgs = [{"role": "user", "content": "Translate the text to Yoruba:  Good morning, how are you?"}]
enc = tok.apply_chat_template(msgs, add_generation_prompt=True, return_tensors="pt", return_dict=True).to("cuda")
with torch.no_grad():
    out = model.generate(**enc, max_new_tokens=256, do_sample=False)
raw = tok.decode(out[0, enc["input_ids"].shape[1]:], skip_special_tokens=True)   # "<yor> ..."
print(re.sub(r"^<[a-z]{3}>\s*", "", raw.strip()))

Training summary

  1. Base model. google/gemma-3-1b-it, with 125 single-token language tags <iso> (for example <yor>, <ibo>, <nso>).
  2. SFT on English to African-language parallel and instruction data.
  3. RL (GRPO) with a reference-free quality-estimation reward.

Full details are on the card of the final model.

Intended use and limitations

Use: archive of earlier checkpoints, kept for reproduction. Use BeardedMonster/gemma-3-1b-translate-best for translation.

  • English to African language only. Other source languages and reverse directions are not trained.
  • Twenty tags have no SFT data and cannot be translated.
  • Weak languages. Several low-resource languages (for example Kanuri, Tamasheq, Dinka, the Arabic dialects) give rough drafts at best.
  • Short inputs. Translate one sentence or a short paragraph at a time; split long documents into sentences.
  • Automatic metrics only. No human evaluation was done. Outputs are machine translations without human verification; review them before use where errors matter (for example medical, legal or safety-critical text).

Citation and licence

Paper in preparation

@misc{Gemma_mt_2026,
  title  = {English-to-African-language translation with small Gemma-3 models: supervised fine-tuning and reinforcement learning},
  author = {{AletheiaAI}},
  year   = {2026},
}

Licence. license: other (commercial-use-with-attribution). Commercial and non-commercial use, modification and redistribution are allowed, with required attribution: credit "Translation models by Aletheia AI" and cite the accompanying paper. These weights are a derivative of Google Gemma 3, so the Gemma Terms of Use and Prohibited Use Policy also apply.

Downloads last month

-

Downloads are not tracked for this model. How to track
Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support

Model tree for BeardedMonster/gemma-3-1b-translate-rl

Finetuned
(588)
this model