Reza2kn/Motarjem-v0.1

Motarjem v0.1

Motarjem v0.1 is a compact 108M-parameter, bidirectional English–Persian translation model based on Falcon-H1-Tiny-Multilingual-100M-Instruct.

This first release is a research checkpoint trained primarily on aligned English and Iranian Persian Wikipedia articles. It is strongest on Wikipedia-like prose. General-domain translation remains a work in progress; see Evaluation and Limitations.

Usage

Motarjem uses three special control tokens. The prompt format is:

<|end_of_text|><|translate|><|source-language|>SOURCE<|target-language|>

Use <|en|> for English and <|fa|> for Persian. Generation should be greedy for the evaluated behavior.

import torch
from transformers import AutoModelForCausalLM, AutoTokenizer

model_id = "Reza2kn/Motarjem-v0.1"
device = "cuda" if torch.cuda.is_available() else "cpu"
dtype = torch.bfloat16 if device == "cuda" else torch.float32

tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(
    model_id,
    dtype=dtype,
    attn_implementation="sdpa",
).to(device).eval()


def translate(text: str, source_language: str, target_language: str) -> str:
    if (source_language, target_language) not in {("en", "fa"), ("fa", "en")}:
        raise ValueError("Supported directions are en->fa and fa->en")

    prompt_ids = [
        tokenizer.bos_token_id,
        tokenizer.convert_tokens_to_ids("<|translate|>"),
        tokenizer.convert_tokens_to_ids(f"<|{source_language}|>"),
        *tokenizer.encode(text, add_special_tokens=False),
        tokenizer.convert_tokens_to_ids(f"<|{target_language}|>"),
    ]
    input_ids = torch.tensor([prompt_ids], device=device)

    with torch.inference_mode():
        output_ids = model.generate(
            input_ids=input_ids,
            attention_mask=torch.ones_like(input_ids),
            do_sample=False,
            max_new_tokens=256,
            eos_token_id=tokenizer.eos_token_id,
            pad_token_id=tokenizer.pad_token_id,
            use_cache=True,
        )

    return tokenizer.decode(
        output_ids[0, input_ids.shape[1]:],
        skip_special_tokens=True,
        clean_up_tokenization_spaces=False,
    ).strip()


print(translate("The weather is beautiful today.", "en", "fa"))
print(translate("امروز هوا خیلی خوب است.", "fa", "en"))

Requires a Transformers release with Falcon-H1 support; this checkpoint was trained and evaluated with transformers==4.57.1.

Training

• Base model: tiiuae/Falcon-H1-Tiny-Multilingual-100M-Instruct
• Parameters after adding direction tokens: 107,909,824
• Dataset: Reza2kn/Wikipedia-FA-EN-DeepSeek-V4-Flash-0731
• Pinned dataset revision: 594f3c54d06f2025a04b393cdac3d640e20f432e
• Training split: 118,822 aligned article pairs
• Frozen holdout: 1,178 article pairs selected deterministically by page ID
• Continued pretraining corpus: 114,249,193 tokens
• Bidirectional supervised corpus: 194,405,474 tokens across 2,622,774 aligned examples
• Maximum training sequence length: 1,024 tokens
• Precision: BF16

The stock Falcon tokenizer was retained and extended with <|translate|>, <|en|>, and <|fa|>. A matched custom Persian-tokenizer pilot was more token-efficient but substantially less accurate, so it was not used for this release.

Evaluation

All scores below use greedy decoding, batch size 128, and a maximum of 256 generated tokens.

Suite	Direction	Rows	BLEU	chrF2	TER	Hit token cap
WMT24++	English → Persian	960	3.93	21.87	114.93	13.75%
WMT24++	Persian → English	960	8.03	32.00	103.74	3.75%
Frozen Wikipedia holdout	English → Persian	1,150	41.34	57.16	54.73	2.43%
Frozen Wikipedia holdout	Persian → English	1,150	41.52	59.49	56.21	1.74%

The frozen evaluation uses the 1,150 cleanly materialized rows from the deterministic 1,178-article holdout allocation.

Compared with the untuned official 100M checkpoint on this project's WMT24++ split, Motarjem v0.1 improved chrF2 from 4.13 to 21.87 for English→Persian and from 10.70 to 32.00 for Persian→English. Results from different model families may use different decoding contracts and should not be treated as a strict architecture comparison.

See evaluation/FINAL_EVALUATION_REPORT.md for the full summary and evaluation/failure-analysis.json for categorized failure counts.

Limitations

• This is an experimental 108M-parameter checkpoint, not a production translation system.
• It is strongly specialized toward Wikipedia-style prose and generalizes poorly to news, social media, chat, short fragments, usernames, and irregular formatting.
• On WMT24++, English→Persian produced repetition loops on 13.85% of rows, wrong-language output on 7.08%, and exact source copies on 3.85%. Persian→English was more reliable but still exhibited repetition and number-preservation failures.
• Tables, markup, URLs, named entities, and numbers can be brittle.
• Long inputs were not validated near the architecture's advertised maximum context. Training used sequences up to 1,024 tokens, and quality degrades on longer held-out inputs.
• Machine-generated Wikipedia translations may contain errors or stylistic artifacts inherited from the training data.

Always evaluate on your own domain and retain human review for consequential translations.

Provenance

• Final model.safetensors SHA-256: 9ac54ec2c56240ba57ce7e3be0251d1d86acf1ebd125e9381500e30bc1661eb0
• The machine-readable training receipt is included as training_receipt.json.
• The base model is distributed under the Falcon LLM License; this derivative checkpoint follows that license.

Citation

If you use the Falcon-H1-Tiny family, please cite its creators:

@misc{falcon_h1_tiny,
  title  = {Falcon-H1-Tiny: A series of extremely small, yet powerful language models redefining capabilities at small scale},
  author = {Falcon-LLM Team},
  year   = {2026}
}