Aladdin-FTI: Multidialect Arabic LLM

• Aladdin-FTI is a fine-tuned language model built to handle Arabic's pluricentric and diglossic nature, supporting translations and generation across multiple dialects and languages.
• It employs a joint MT+GEN training strategy that balances semantic fidelity (ChrF++) and dialect authenticity (Macro ADI2), outperforming MT-only and GEN-only approaches.
• The system addresses key challenges in Arabic NLP such as dialect variation, non-standard orthography, and the MSA-dialect asymmetry, advancing multidialectal language processing.

Aladdin-FTI is a set of instruction fine-tuned LLMs developed by the University of Geneva’s Faculty of Translation and Interpreting for the AMIYA shared task, “Arabic Modeling In Your Accent.” It is designed as a single model family that both generates fluent text in multiple Arabic dialects and translates between these dialects, Modern Standard Arabic (MSA), and English, while explicitly treating Arabic as a pluricentric, diglossic language rather than “just MSA” (Mutal et al., 18 Feb 2026). The system supports Moroccan, Egyptian, Palestinian, Syrian, and Saudi dialects, and its central methodological claim is that joint optimization of machine translation and dialect-conditioned generation improves both semantic fidelity and dialect faithfulness.

1. Definition and task setting

Aladdin-FTI was built for the AL-QASIDA benchmark used in AMIYA at VarDial 2026. The benchmark evaluates dialectal Arabic LLMs along two complementary axes. The first is Diglossia, defined operationally as performance on translation between MSA and dialects, measured with ChrF++. The second is Fidelity, defined in the shared-task setting as the degree to which generated text is dialectal and correctly identifiable as the requested dialect, measured with Macro ADI2, a dialect-classification-based score (Mutal et al., 18 Feb 2026).

The system participates in the closed-data track. In that setting, fine-tuning is restricted to official organizer-provided training corpora together with limited, specified dialectal Arabic resources; no arbitrary external data or in-context tricks are permitted. The intended scope therefore includes both translation subtasks—notably MSA↔DA and DA↔English—and generation subtasks, where prompts request output in a specified dialect.

The immediate motivation is a long-standing asymmetry in Arabic NLP. Most prior work has centered on MSA, while Arabic usage in everyday communication is fundamentally diglossic and dialectally heterogeneous. The paper identifies four obstacles that follow from this: diglossia itself, substantial dialect variation relative both to MSA and to other dialects, non-standard orthography, and scarcity of high-quality parallel and monolingual dialect corpora. A recurrent failure mode of broadly pretrained LLMs in this setting is re-normalization of dialectal inputs or outputs back toward MSA.

2. The three desiderata: fidelity, diglossia, and multidialectal generation

The title’s “three wishes” correspond to three technical objectives. Fidelity denotes preservation of source meaning, avoidance of hallucinations or semantic drift, and—on translation tasks—high adequacy with acceptable fluency. In the evaluation regime, translation fidelity is proxied by higher ChrF++, whereas for open-ended generation the relevant notion is semantic soundness relative to the prompt.

Diglossia is not treated as a purely sociolinguistic label but as a modeling capability: the model must distinguish MSA from multiple dialects and perform accurate bidirectional MSA↔DA translation. In this operationalization, higher ChrF++ indicates stronger diglossic competence. The paper reports that MT-only fine-tuning yields the highest ChrF++ but comparatively modest dialect fidelity.

Multidialectal generation denotes on-demand generation in multiple dialects while preserving dialect-specific lexical items, morphology, syntax, and orthographic conventions. This capability is evaluated with Macro ADI2, which tests whether outputs are dialectal at all and whether they are classified as the correct target dialect. The paper’s central trade-off is that GEN-only training raises Macro ADI2 but harms ChrF++, whereas MT-only training improves ChrF++ but tends to produce outputs that are too close to MSA or to neutral Arabic. Joint MT+GEN training is presented as the mechanism that balances these competing pressures (Mutal et al., 18 Feb 2026).

A plausible implication is that the system’s contribution is less a new architecture than a new operating point on the fidelity–dialectality frontier. The paper explicitly frames the objective as simultaneous preservation of semantic fidelity across languages and registers and production of dialect-faithful output.

3. Architecture and optimization strategy

Aladdin-FTI uses decoder-only LLMs and does not introduce a sequence-to-sequence architecture, bespoke Arabic tokenizer, or modular dialect-specific heads. The two principal base models are SmolLM3-3B, used as the main AMIYA submission, and Llama-3.1-8B-Instruct, used to study scaling effects. For comparison without fine-tuning, the paper reports results for GPT-OSS-120B and Command R Arabic (Mutal et al., 18 Feb 2026).

Both tuned models are instruction fine-tuned with a causal LM objective over chat-style data. All tasks are formatted as instruction–assistant exchanges, and the loss is computed only on assistant tokens:

$$\mathcal{L}(\mathbf{z}) = - \sum_{t \in \text{assistant tokens}} \log p_\theta(w_t \mid w_{<t})$$

The multi-task formulation combines two data sources: machine translation examples and instruction-based next-token generation examples:

$$\mathcal{L}_{\text{joint}}(\theta) = \lambda \, \mathbb{E}_{\mathbf{z}\sim\mathcal{D}_{\text{MT}}}\big[\mathcal{L}(\mathbf{z})\big] + (1-\lambda)\,\mathbb{E}_{\mathbf{z}\sim\mathcal{D}_{\text{GEN}}}\big[\mathcal{L}(\mathbf{z})\big].$$

Here, $\lambda=1$ corresponds to MT-only training, $\lambda=0$ to GEN-only training, and $\lambda=0.5$ to joint MT+GEN training. The experiments explicitly compare all three settings.

The adaptation regime differs by model. SmolLM3-3B is fully fine-tuned, in bfloat16, with gradient checkpointing and no LoRA. Llama-3.1-8B-Instruct is fine-tuned with LoRA using rank $r=16$, $\alpha=32$, dropout 0.05, applied to attention projections $(q,k,v,o)$ and feed-forward layers. No quantization is used.

Input processing is intentionally light. The models retain their standard subword tokenizers; there is no heavy orthographic normalization because the stated goal is to respect native, non-standard spellings. Pre-processing mainly wraps data into instruction templates with explicit language or dialect labels and constructs generation examples from three-word sentence prefixes. Post-processing consists of direct decoding with controlled sampling hyper-parameters, especially temperature and top-$p$. In this system, dialect control is implemented by natural-language instructions rather than special tokens or constrained decoding.

Training is run for 4 epochs with effective batch size 128, AdamW, cosine learning-rate scheduling with 3% warmup, weight decay 0.01, and gradient clipping at max norm 1.0. The paper reports testing learning rates $$\{2\mathrm{e}{-5}, 3\mathrm{e}{-5}, 5\mathrm{e}{-5}, 6\mathrm{e}{-5}\}$$ and selecting checkpoints by macro-averaged ChrF++ after evaluation every 1,000 steps using deterministic generation with $$\mathcal{L}_{\text{joint}}(\theta) = \lambda \, \mathbb{E}_{\mathbf{z}\sim\mathcal{D}_{\text{MT}}}\big[\mathcal{L}(\mathbf{z})\big] + (1-\lambda)\,\mathbb{E}_{\mathbf{z}\sim\mathcal{D}_{\text{GEN}}}\big[\mathcal{L}(\mathbf{z})\big].$$0 and top-$$\mathcal{L}_{\text{joint}}(\theta) = \lambda \, \mathbb{E}_{\mathbf{z}\sim\mathcal{D}_{\text{MT}}}\big[\mathcal{L}(\mathbf{z})\big] + (1-\lambda)\,\mathbb{E}_{\mathbf{z}\sim\mathcal{D}_{\text{GEN}}}\big[\mathcal{L}(\mathbf{z})\big].$$1 (Mutal et al., 18 Feb 2026).

4. Language coverage, prompting, and training data

The system covers five primary dialects in evaluation—Moroccan, Egyptian, Palestinian, Syrian, and Saudi—together with MSA and English. Jordanian also appears in training sources, although the primary evaluation focuses on the five main regional varieties. Supported capabilities include monolingual generation in target dialects, MSA↔DA translation, DA↔English translation where parallel corpora exist, and implicitly DA↔DA translation through flexible instruction templates and multidialect training data (Mutal et al., 18 Feb 2026).

The control mechanism is entirely instruction based. For translation, the template specifies source and target varieties directly, as in “Translate from English into Egyptian Arabic. Output only the translation.” For generation, the model receives either an English instruction such as “Complete the sentence in Egyptian Arabic…” or a dialectal instruction written in the target dialect itself. This dual prompting regime teaches the model both to interpret dialect labels in English and to respond to instructions phrased in the dialect being requested.

The training corpus is split into bilingual data for $$\mathcal{L}_{\text{joint}}(\theta) = \lambda \, \mathbb{E}_{\mathbf{z}\sim\mathcal{D}_{\text{MT}}}\big[\mathcal{L}(\mathbf{z})\big] + (1-\lambda)\,\mathbb{E}_{\mathbf{z}\sim\mathcal{D}_{\text{GEN}}}\big[\mathcal{L}(\mathbf{z})\big].$$2 and monolingual data for $$\mathcal{L}_{\text{joint}}(\theta) = \lambda \, \mathbb{E}_{\mathbf{z}\sim\mathcal{D}_{\text{MT}}}\big[\mathcal{L}(\mathbf{z})\big] + (1-\lambda)\,\mathbb{E}_{\mathbf{z}\sim\mathcal{D}_{\text{GEN}}}\big[\mathcal{L}(\mathbf{z})\big].$$3. The bilingual side includes SauDial, Casablanca, JODA, UFAL parallel corpus, DODA, Atlas training sets, and material related to MADAR26. The monolingual side includes SDC corpus, Saudi Tweets Corpus, Shami corpus, MASC, Goud, EDGAD, EDC, ASR-EgAr, and the MADAR training set, which is used specifically to build sentence-completion tasks in many dialects.

The generation objective is built around sentence completion from a three-word prefix rather than unconstrained free generation. This design matters because it makes the training signal explicitly conditional on both a partial surface form and a requested dialect. The paper argues that, because many continuations are valid, the objective encourages the model to internalize distributional dialect properties rather than memorize fixed outputs.

5. Evaluation protocol and empirical profile

Evaluation follows AL-QASIDA. Translation-oriented diglossia tasks use ChrF++ as a corpus-level metric over test sets. Generation-oriented fidelity tasks use Macro ADI2, a macro-averaged dialect identification score over dialects and datasets. Generation prompts are drawn from Okapi prompts, ShareGPT prompts, and human-curated prompts (Marchisio et al. 2024); parallel evaluation data include MADAR26, FLORES200 dev subsets for 5 regional dialects, and the HABIBI song lyrics corpus. The paper reports no human evaluation; all metrics are automatic (Mutal et al., 18 Feb 2026).

The following table summarizes the headline comparison reported for baseline and comparison systems.

Model	Diglossia (ChrF++)	Fidelity (Macro ADI2)
SmolLM3-3B (baseline)	22.23	0.003
+ MT + Instruction (contrastive)	33.65	0.067
Llama-3.1-8B-Instruct (baseline)	32.99	0.065
+ MT + Instruction (primary)	35.09	0.233
Command R Arabic	46.60	0.053
GPT-OSS-120B	47.82	0.237

The task-type ablation is the paper’s main empirical result. MT-only fine-tuning strongly improves ChrF++ but yields only modest and variable gains in Macro ADI2. GEN-only fine-tuning substantially improves Macro ADI2 but significantly degrades ChrF++, indicating semantic drift and poor translation quality. MT+GEN largely preserves the ChrF++ gains of MT-only training while substantially improving Macro ADI2 relative to MT-only. In the paper’s framing, this answers its three research questions directly: MT improves diglossia, GEN improves fidelity but hurts diglossia, and MT+GEN gives the best trade-off.

Scaling is informative but not monotonic. Fine-tuning improves both model sizes. For SmolLM3-3B, joint training raises ChrF++ by more than 11 points relative to baseline. For Llama-3.1-8B-Instruct, joint training raises Macro ADI2 from 0.065 to 0.233. The resulting fidelity score is close to GPT-OSS-120B at 0.237, while Command R Arabic posts stronger ChrF++ but much lower fidelity at 0.053. This suggests that dialect control in this benchmark depends heavily on fine-tuning objective design rather than model size alone.

Decoding settings also matter. Moderate temperatures and top-$$\mathcal{L}_{\text{joint}}(\theta) = \lambda \, \mathbb{E}_{\mathbf{z}\sim\mathcal{D}_{\text{MT}}}\big[\mathcal{L}(\mathbf{z})\big] + (1-\lambda)\,\mathbb{E}_{\mathbf{z}\sim\mathcal{D}_{\text{GEN}}}\big[\mathcal{L}(\mathbf{z})\big].$$4 values produce stable ChrF++ across models, whereas very high temperature with top-$$\mathcal{L}_{\text{joint}}(\theta) = \lambda \, \mathbb{E}_{\mathbf{z}\sim\mathcal{D}_{\text{MT}}}\big[\mathcal{L}(\mathbf{z})\big] + (1-\lambda)\,\mathbb{E}_{\mathbf{z}\sim\mathcal{D}_{\text{GEN}}}\big[\mathcal{L}(\mathbf{z})\big].$$5 can sharply reduce ChrF++ for small models; the paper gives SmolLM3-3B diglossia 9.78 as an example under that regime. GPT-OSS-120B is reported to be more robust across decoding settings.

6. Error profile, limitations, and place in the literature

Qualitative examples show that Aladdin-FTI can produce natural colloquial translation and dialect-conditioned continuations. The paper highlights an EN→Egyptian example, “I feel chilly and my stomach hurts badly” → “حاسس إني بردان ومعدتي واجعاني جامد.” It also reports that the model can interpret dialectal instructions such as “كمّل الجملة...” and continue coherently in the requested variety (Mutal et al., 18 Feb 2026).

The paper does not provide a formal human-annotated error taxonomy, but it does identify several recurring failure modes from metric behavior and broader context. These include dialect confusion, where one dialect bleeds into another; MSA leakage, especially under MT-only training; semantic drift under GEN-only training; and vulnerability to non-standard spelling variability beyond what appears in training. The absence of human evaluation remains a central limitation, particularly because Macro ADI2 rewards recognizability by a dialect identifier rather than directly measuring adequacy or naturalness.

The system’s broader significance lies in its treatment of Arabic as a family of related varieties with distinct registers rather than as a monolithic MSA-centric object. In the paper’s self-positioning, Aladdin-FTI sits at the intersection of Arabic dialect MT, dialectal text generation with LLMs, and multidialectal or pluricentric language modeling. Earlier dialect MT systems are described as prone to dialect normalization; dialectal generation systems can improve controllability but may harm adequacy. Aladdin-FTI’s main methodological claim is that joint MT+GEN instruction fine-tuning addresses both tendencies within a single unified LLM.

The released resources reinforce that positioning. The authors provide code at https://github.com/drvenabili/mtfinetune_amiya and model artifacts at https://hf.co/collections/unige-fti/aladdin-fti-amiya. The release includes training scripts for both objectives, instruction templates, fine-tuning and LoRA hyper-parameters, and an evaluation pipeline aligned with AL-QASIDA. This makes the system reproducible within the closed-track data regime and, as the paper notes, extensible to additional dialects by adding appropriately labeled MT and monolingual data.