Pith. sign in

REVIEW 4 major objections 5 minor 26 references

GemMaroc: Unlocking Darija Proficiency in LLMs with Minimal Data

T0 review · 4 major / 5 minor · reviewed 2026-08-07 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read The paper claims that a few thousand translated, reasoning-heavy instructions are enough to give Gemma models Moroccan Darija proficiency without sacrificing English reasoning, at a fraction of the usual compute.

desk verdict Useful cheap recipe and a real new Darija data resource, but the fluency claim outruns the evidence because the translation pipeline leans MSA and the benchmarks may not separate dialect from standard Arabic. read the letter →

arxiv 2505.17082 v1 pith:LYKD3BVE submitted 2025-05-20 cs.CL cs.AI

Abderrahman Skiredj , Ferdaous Azhari , Houdaifa Atou , Nouamane Tazi , Ismail Berrada This is my paper · ORCID
classification cs.CLcs.AI
keywords DarijaMoroccanArabicinstructiontuninglow-resourcelanguagemodelingLoRAminimaldatacross-lingualreasoningGreenAI
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

The paper claims that Moroccan Arabic (Darija) can be elicited from an off-the-shelf Gemma model using far less data and compute than prior efforts, provided the instruction data is carefully selected and reasoning-dense. It translates three compact instruction suites into Darija, keeps about a fifth in English, and fine-tunes Gemma 3 with LoRA. On the 4B model, a few thousand multi-turn instructions raise DarijaMMLU (a 22,027-question multiple-choice benchmark) by 9.9 points without hurting English; on the 27B model, the recipe matches the strongest published Darija score on DarijaMMLU and beats it by 12.1 points on DarijaHellaSwag while keeping GSM8K at 84.2%. The entire 27B run takes 48 GPU-hours, under 2% of the energy reported for the main baseline. If right, this shows dialect competence and reasoning are not in tension, and inclusive language models need not be expensive.

What carries the argument

The load-bearing object is a small, information-dense instruction suite built by translating three datasets into Darija: LIMA-1K is one thousand curated single-turn prompts; DEITA-6K is six thousand multi-turn prompts selected for complexity, quality, and diversity; and the TULU slice is a roughly 46,000-example reasoning-heavy subset rich in mathematics, coding, and science. The translation is done automatically with Gemini 2.0 Flash under a strict prompt that keeps code, LaTeX, and English technical terms unchanged, and about 20-30% of each set is left in English to prevent catastrophic forgetting. These instructions are applied through LoRA adapters (low-rank updates that leave the pretrained weights frozen) on Gemma 3-4B and Gemma 3-27B. The mechanism works by giving the model just enough dialect-specific style and reasoning examples to unlock Darija knowledge already latent in pretraining, while the remaining English data preserves cross-lingual reasoning.

What would settle it

Ask native Darija speakers to judge a random sample of GemMaroc-27B's responses for dialectal authenticity and correctness, or retrain the same recipe on human-authored Darija instructions and compare; a systematic MSA or translationese pattern, or a human-data model that fails to reproduce the gains, would show the reported scores do not reflect genuine Darija learning.

Watch

Extended reading notes

Core claim

The central discovery is that quality, not quantity, carries Darija alignment. A 4B Gemma model tuned with roughly 5,000 curated Darija instructions rises from 32.8% to 42.7% on DarijaMMLU, and adding a 46,000-example reasoning-dense slice pushes it to 47.5% with no English regression. Scaling the same recipe to Gemma 3-27B yields GemMaroc-27B, which ties Atlas-Chat on DarijaMMLU at 61.6% and surpasses all prior models on DarijaHellaSwag at 60.5% versus Atlas-Chat's 48.4%, while retaining 84.2% on English GSM8K. All of this is done with LoRA adapters in 48 GPU-hours, an order of magnitude less supervision and roughly 2% of the energy footprint reported for comparable tuning. The paper takes this as evidence that reasoning-centric instructions are the most data-efficient lever for surfacing latent Darija competence in a pretrained multilingual backbone.

Load-bearing premise

The automatic Gemini 2.0 Flash translations, which the authors sampled but did not verify exhaustively, must be natural Darija rather than Modern Standard Arabic with translation artifacts; if they are not, the benchmark gains would not demonstrate real dialect competence.

Editorial extensions

If this is right

  • A few thousand well-chosen Darija instructions are enough to move a 4B model about 10 points on DarijaMMLU, placing the tipping point between 1,000 and 6,000 examples.
  • Adding reasoning-dense translated data improves both Darija benchmarks and English maths and commonsense, so dialect tuning need not trade away reasoning ability.
  • A 27B LoRA-tuned model can match a model trained on roughly 458,000 instructions on DarijaMMLU and surpass it by 12.1 points on DarijaHellaSwag, using about one-tenth the supervision.
  • Forty-eight GPU-hours (roughly 26-32 kWh and under $100 of cloud cost) is enough to produce a deployable 27B Darija model, lowering the entry barrier for public-sector use.
  • Keeping 20-30% of training data in English preserves world knowledge and mathematical reasoning, with GSM8K holding at 84.2%.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • If the machine-translated data is the bottleneck, replacing it with human-written or human-verified Darija could widen the gap further; the paper's own check is a sample review, not a full audit.
  • The same quality-over-quantity recipe may transfer to other under-represented dialects, but only if the base model already has enough latent exposure to that variety during pretraining.
  • Because the Darija benchmarks are themselves machine-rendered, the reported gains may overstate everyday conversational fluency until native-speaker evaluation of GemMaroc's open-ended responses is done.
  • Keeping the original SentencePiece tokenizer, which was not built for Darija's orthography, likely explains the weaker sentiment and summarization scores, making a script-aware tokenizer a natural next step.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

4 major / 5 minor

Summary. The paper proposes a minimal-data, Green-AI recipe for surfacing Moroccan Arabic (Darija) in Gemma models. It translates three existing English instruction suites (LIMA 1K, DEITA 6K, and a 50K TULU subset) into Darija via Gemini 2.0 Flash, keeps 20% of English examples, and LoRA-tunes Gemma 3-4B and Gemma 3-27B. The authors report that a 4B model trained on roughly 5K translated DEITA instructions raises DarijaMMLU from 32.8% to 42.7%, that adding the TULU slice pushes it to 47.5%, and that the 27B model reaches 61.6% on DarijaMMLU and 60.5% on DarijaHellaSwag, with an estimated total cost of 48 GPU-hours. The paper also releases code, data, and checkpoints.

Significance. If the Darija-competence claim holds, the result is practically and environmentally significant: it suggests that a small, reasoning-dense, translated instruction set can unlock a low-resource dialect with a tiny fraction of the compute used by prior Arabic-centric models. The paper's strengths are its open assets, its transparent (though internally inconsistent) energy accounting, and the fact that its recipe is borrowed from published data-selection methods rather than fitted to the target benchmark, so I do not see a circularity problem. The main risk is construct validity: because the translation protocol explicitly falls back to Modern Standard Arabic for formal, scientific, and uncertain content, the reported benchmark gains may reflect an MSA/translationese mixture rather than genuine Darija fluency. This issue is acknowledged in the paper's own Limitations section but is not resolved by the current evidence.

major comments (4)
  1. [§3.1, Appendix A.1.1, Limitations] The translation protocol instructs the model to prefer Modern Standard Arabic when uncertain, to use MSA for formal or scientific content, and to use MSA whenever a Darija rendering would sound awkward, unclear, or artificial. Since the TULU slice is composed largely of mathematics, coding, and scientific prompts, a substantial part of the 'Darija' training set is likely MSA or code-switched translationese. The evaluation benchmarks (DarijaMMLU and DarijaHellaSwag) are themselves translated academic and commonsense tests, so they cannot distinguish Darija from an MSA-leaning translated register. The Limitations section concedes that the data relies on machine-translated Darija samples with limited verification, and the manual verification described in §3.1 was prompt refinement rather than a systematic fluency or dialect-identity check. The central claim that GemMaroc 'surfaces fluent Darija' therefore requires direct evidence: native-speaker judgments of model outputs, human annotation of a random sample of the training translations, or a Darija-specific benchmark with MSA-distractor controls. Without such evidence, the reported gains are equally consistent with the model having learned a formal-Arabic/translationese mixture.
  2. [Table 2, §4.3] The leaderboard lacks an untuned Gemma-3-27B row. The paper claims that GemMaroc-27B 'retains Gemma-27B's strong maths and general-reasoning ability' and is 'within 12 pp of the untuned backbone' on GSM8K, but no Gemma-3-27B baseline appears anywhere in Table 2 or the text. Since the 4B rows include the untuned Gemma3-4b as a baseline, the omission for the 27B model is likely an oversight but it directly undermines the reasoning-retention claim. Add the untuned Gemma-3-27B scores for all English benchmarks (and ideally for the Darija benchmarks as well), and report the retention deltas explicitly.
  3. [Abstract, §4.3, Appendix A.3] The energy-cost claim is internally inconsistent. The abstract and the introduction state that the entire model is trained in 48 GPU-hours and consumes about 26 kWh, while Appendix A.3 Table 4 lists a total of 58 GPU-hours (10 A100 GPU-hours for the 4B ablations plus 48 H100 GPU-hours for the 27B run) and computes 32 kWh total and about 13 kg CO2e. The discussion in §4.3 also gives 26 kWh and 10 kg CO2e using a 0.54 kW/GPU draw, whereas the appendix uses 0.5 kW for H100 and a 0.9 utilization factor. Since the Green-AI contribution is a load-bearing part of the paper, these numbers must be reconciled and the protocol stated once, clearly.
  4. [Table 2, §4.3] The discussion states that adding the TULU slice keeps 'the MMLU drop below 3 pp,' but Table 2 shows the opposite: MMLU rises from 51.1 (untuned Gemma3-4b) and 51.35 (DEITA) to 54.14 (TULU). The sentence is contradicted by the paper's own table and should be corrected or the intended comparison clarified (e.g., if it refers to a different evaluation run).
minor comments (5)
  1. [§1, Contributions] The contribution bullet claims GemMaroc-27B is 'the first open Darija LLM to reach 61.6% DarijaMMLU,' but Table 2 shows Atlas-Chat-27B already scores 61.95% on DarijaMMLU. The novelty should be phrased as matching the prior best and setting a new high-water mark on DarijaHellaSwag.
  2. [§3.1, Table 1] The text repeatedly calls the DEITA corpus 'DEITA 6K' and says it contains 6,000 prompts, but Table 1 lists 5,000 total samples (3,700 translated + 1,300 English). Please clarify whether this reflects post-translation filtering or a naming inconsistency, and apply the same clarity to the TULU count (50K vs. 46K).
  3. [Table 2] All reported scores are single-run results with no error bars or significance tests. Given that several headline comparisons are small (e.g., 61.61 vs. 61.95 on DarijaMMLU), report multiple seeds or confidence intervals, or temper the 'matches' and 'retains' language accordingly.
  4. [Appendix A.2] The illustrative Darija examples (Figures 1-7) appear as empty placeholders in the submitted text; if the figures are part of the intended submission, they should be included.
  5. [§2.1] The paragraph on Command R+ / Command A is only loosely connected to the Darija-focused narrative and could be trimmed or explicitly tied to the paper's methodology.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: the paper reports empirical fine-tuning results against externally constructed benchmarks, with no fitted-input-as-prediction or self-citation load-bearing steps.

full rationale

GemMaroc is an empirical study: it translates existing instruction suites (LIMA, DEITA, TULU), fine-tunes Gemma checkpoints with LoRA, and evaluates on externally published benchmarks (DarijaMMLU, DarijaHellaSwag, GSM8K, MMLU, etc., from Shang et al. 2025, Hendrycks et al. 2021, Cobbe et al. 2021). The central claims—that roughly 6k translated instructions raise DarijaMMLU by +9.9 pp on a 4B backbone and that the 27B model reaches 61.6% DarijaMMLU / 60.5% DarijaHellaSwag—are reported measurements, not derivations from fitted parameters. Hyperparameters are stated to be 'inspired by the settings recommended in Zhou et al. (2023); Liu et al. (2024); Chen et al. (2025)', and no parameter is fitted to the evaluation benchmarks. The paper does not invoke a uniqueness theorem, and its citations to prior work (LIMA, DEITA, MIG, Atlas-Chat) are external evidence, not self-citations by the present authors. The only notable weakness is validity-related, not circular: the Limitations section concedes the data 'relies on machine-translated Darija samples with limited verification', and the Appendix translation prompt directs the translator to prefer MSA when Darija 'would sound awkward, unclear, or artificial'; this threatens whether benchmark gains reflect genuine Darija competence, but it is not a case of a prediction being equivalent to its input by construction. Under the stated criteria, no circular step is present.

Assumptions & free parameters 5 free parameters · 5 assumptions · 0 invented entities

The algorithm uses standard SFT with LoRA and introduces no mathematical axioms or invented entities. The load-bearing assumptions are empirical: that the base models have latent Darija, that machine translation is faithful, that the benchmarks are valid and uncontaminated, and that the chosen hyperparameters and English-retention ratio are adequate. None of these assumptions is independently verified beyond author sampling and borrowed practice from cited prior work.

free parameters (5)
  • LoRA rank and alpha = r=32, alpha=64 (4B); r=16, alpha=32 (27B)
    Hand-chosen capacity settings, not swept or fitted to Darija benchmarks. Borrowed from cited minimal-SFT recipes.
  • Learning rate per dataset = 1e-4 (TULU); 4e-4 (DEITA, LIMA)
    Chosen following Zhou et al., Liu et al. and Chen et al.; no reported tuning.
  • Epochs per dataset = 3 (TULU); 6 (DEITA); 15 (LIMA)
    Set per dataset without ablations; different epoch counts make comparisons across corpora somewhat confounded.
  • English retention ratio = 20% claimed; 26-30% in Table 1
    Design choice to mitigate catastrophic forgetting; no ablation varying the ratio, and realized ratios differ from the stated 20%.
  • Token-length cutoff = 2,048 tokens
    Samples above this length are dropped; this changes DEITA from 6k to 5k and affects which TULU examples are kept.
assumptions (5)
  • domain assumption Gemma 3 checkpoints contain latent Darija knowledge that can be surfaced by SFT.
    The mechanism behind the minimal-data claim, stated in the Introduction and Discussion as 'Darija knowledge acquired during pre-training merely needed the right key.' Not directly measured.
  • domain assumption Automatic Gemini 2.0 Flash translation into Darija is faithful enough for effective SFT.
    Section 3.1 describes the translation pipeline; only sampled author review is reported, and the Limitations section admits 'limited verification.'
  • domain assumption DarijaMMLU and DarijaHellaSwag are valid measures of Darija competence and are not contaminated by the training data.
    Section 3.3 adopts benchmarks from Shang et al.; no contamination or leakage analysis is provided.
  • domain assumption Retaining roughly a fifth to a quarter of samples in English prevents catastrophic forgetting and preserves cross-lingual reasoning.
    Section 3.1 cites prior work for this practice; the paper performs no ablation on the retention ratio.
  • domain assumption Filtering out non-English instructions before translation avoids translation artifacts from meta-prompts.
    Section 3.1 filtering step is plausible but not validated empirically.

how reviews work

0 comments
Cite this review

Pith. "Pith review of GemMaroc: Unlocking Darija Proficiency in LLMs with Minimal Data." pith.science (2026). https://pith.science/paper/LYKD3BVE

@misc{pith2026250517082,
  author       = {Pith},
  title        = {Pith review of: GemMaroc: Unlocking Darija Proficiency in LLMs with Minimal Data},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/LYKD3BVE}},
  note         = {Machine review of arXiv:2505.17082}
}
read the original abstract

Open-source large language models (LLMs) still marginalise Moroccan Arabic (Darija), forcing practitioners either to bolt on heavyweight Arabic adapters or to sacrifice the very reasoning skills that make LLMs useful. We show that a rigorously quality-over-quantity alignment strategy can surface fluent Darija while safeguarding the backbone s cross-lingual reasoning at a sliver of the usual compute. We translate three compact instruction suites LIMA 1 K, DEITA 6 K and TULU 50 K into Darija, preserve 20 of the English originals, and add mathematics, coding and scientific prompts. A LoRA-tuned Gemma 3-4B trained on 5 K mixed instructions lifts DarijaMMLU from 32.8 to 42.7 ; adding the reasoning-dense TULU portion pushes it to 47.5 with no English regression. Scaling the identical recipe to Gemma 3-27B produces GemMaroc-27B, which matches Atlas-Chat on DarijaMMLU (61.6 ) and leaps ahead on Darija commonsense, scoring 60.5 on HellaSwag versus Atlas-Chat s 48.4 . Crucially, GemMaroc retains Gemma-27B s strong maths and general-reasoning ability, showing only minimal movement on GSM8K and English benchmarks. The entire model is trained in just 48 GPU.h, underscoring a Green AI pathway to inclusive, sustainable language technology. We release code, data and checkpoints to spur Darija-centric applications in education, public services and everyday digital interaction.

Figures

Figures reproduced from arXiv: 2505.17082 by the authors.

Figure 1
Figure 1. Illustrative example 1 from Darija Lima dataset. [PITH_FULL_IMAGE:figures/full_fig_p018_1.png] view at source ↗
Figure 2
Figure 2. Illustrative example 2 from Darija Lima dataset. [PITH_FULL_IMAGE:figures/full_fig_p019_2.png] view at source ↗
Figure 3
Figure 3. Round 1 of an Illustrative example from Darija Deita dataset. [PITH_FULL_IMAGE:figures/full_fig_p020_3.png] view at source ↗
Figures from the paper (4 more)
Figure 4
Figure 4. Figure 4: Round 2 of an Illustrative example from Darija Deita dataset. [PITH_FULL_IMAGE:figures/full_fig_p020_4.png]
Figure 5
Figure 5. Figure 5: Round 3 of an Illustrative example from Darija Deita dataset. [PITH_FULL_IMAGE:figures/full_fig_p021_5.png]
Figure 6
Figure 6. Figure 6: Illustrative example 1 from Darija TULU dataset. [PITH_FULL_IMAGE:figures/full_fig_p022_6.png]
Figure 7
Figure 7. Figure 7: Illustrative example 2 from Darija TULU dataset. [PITH_FULL_IMAGE:figures/full_fig_p023_7.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

26 extracted references · 24 canonical work pages

  1. [1]

    Translate comments inside code blocks to Dar- ija (Arabic script)

    Code blocks: Keep all programming code (Python, JS, etc.) unchanged. Translate comments inside code blocks to Dar- ija (Arabic script)

    work page

  2. [2]

    Do not alter them in any way

    Math formulas and LaTeX: Preserve all mathematical notation and LaTeX equations. Do not alter them in any way

    work page

  3. [3]

    Alexandre Lacoste, Alexandra Luccioni, Victor Schmidt, and Thomas Dandres

    Muri: High-quality instruction tuning datasets for low-resource languages via reverse instructions. Alexandre Lacoste, Alexandra Luccioni, Victor Schmidt, and Thomas Dandres. 2019. Quantifying the carbon emissions of machine learning. arXiv preprint arXiv:1910.09700. Huiyuan Lai and Malvina Nissim. 2024. mCoT: Multi- lingual instruction tuning for reasoni...

    arXiv 2019

  4. [4]

    In Pro- ceedings of the First Workshop on Language Mod- els for Low-Resource Languages, pages 9–30, Abu Dhabi, United Arab Emirates

    Atlas-chat: Adapting large language models for low-resource moroccan arabic dialect. In Pro- ceedings of the First Workshop on Language Mod- els for Low-Resource Languages, pages 9–30, Abu Dhabi, United Arab Emirates. Association for Com- putational Linguistics. Emma Strubell, Ananya Ganesh, and Andrew McCal- lum. 2019. Energy and policy considerations fo...

    arXiv 2019

  5. [5]

    In Proceedings of the 62nd Annual Meeting of the Association for Com- putational Linguistics (Volume 1: Long Papers) , pages 7025–7046, Bangkok, Thailand

    PLUG: Leveraging pivot language in cross- lingual instruction tuning. In Proceedings of the 62nd Annual Meeting of the Association for Com- putational Linguistics (Volume 1: Long Papers) , pages 7025–7046, Bangkok, Thailand. Association for Computational Linguistics. Chunting Zhou, Pengfei Liu, Puxin Xu, Srini Iyer, Jiao Sun, Yuning Mao, Xuezhe Ma, Avia E...

    work page 2023

  6. [6]

    Strong growth potential with community engagement. Example 3: Data Warehouse Simplifications Human: explain data warehouse architecture in easy language Assistant: A data warehouse collects data from various sources, transforms it, and stores it centrally. Main components:

    work page

  7. [8]

    é ªJ “” for “formula

    Technical Terms: Keep technical/scientific terms in English un- less they have a known Arabic equivalent. For example: • Use “ é ªJ “” for “formula” • Use “ éËXA ªÓ” for “equation” • Use “ éË@X ” for “function” • But do not translate terms like “API”, “eigenvalue”, “html”, etc

    work page

  8. [9]

    TRANSLATE into Moroccan Darija (Arabic script): • Regular sentences, explanations, and natural narrative text

    Proper Names: Do not translate names of people, tools, or software (e.g., Python, Newton, GitHub, etc.). TRANSLATE into Moroccan Darija (Arabic script): • Regular sentences, explanations, and natural narrative text. • Comments inside code blocks (e.g., # comment here ). • Use a natural mix of Moroccan Darija and Modern Standard Arabic where appropriate: –...

    work page 2023

Show all 26 references
  1. [10]

    Flutter Entertainment

  2. [11]

    Bet365 Human: competitors who are also expanding into esports, game development Assistant: Companies include:

  3. [12]

    Assistant: As of my knowledge cutoff, Unikrn has not acquired any gaming studios

    Unikrn Human: what studios has Unikrn acquired? Furthermore, elaborate on äcquired¨. Assistant: As of my knowledge cutoff, Unikrn has not acquired any gaming studios. Explanation of acquisitions:

  4. [13]

    Tencent/Riot Games: Full acquisition over time

  5. [14]

    Tencent/Supercell: Majority stake through a consortium

  6. [15]

    Example 2: IP Ownership and Business Model Human: I have ownership of the IP of video games created by S2 Games excluding Heroes of Newerth

    Activision Blizzard: Complete acquisition of Vivendi Games. Example 2: IP Ownership and Business Model Human: I have ownership of the IP of video games created by S2 Games excluding Heroes of Newerth... What is the appeal of this, to any investor? Assistant: Key appeals include:

  7. [16]

    Nostalgia and a loyal player base

  8. [17]

    Unique IP and lore for expansion

  9. [18]

    Use of AI-generated art and modern features

  10. [19]

    Platform potential for indie developers

  11. [20]

    Revenue-sharing model incentives

  12. [22]

    Summary: DEITA-6k demonstrates that quality, complexity, and diversity can be engineered simultaneously to produce a highly effective dataset for alignment

    Front-End Tools Example 4: Contracted Chief Brand Officer’s Role Human: Generate a description of a Contracted Chief Brand Officer’s role Assistant: A Contracted Chief Brand Officer becomes an integral part of your executive team, leading brand strategy and development, workin...

    2025

  13. [23]

    Calculate the total displacement vector d: Given vectors: a = (3, 2, −1), b = (−2, 4, 5), c = (1, −3, 2) The total displacement vector is: d = a + b + c Compute each component: dx = 3 + (−2) + 1 = 2, dy = 2 + 4 + (−3) = 3, dz = −1 + 5 + 2 = 6 So, the total displacement vector ...

  14. [24]

    Since only the vertical displacement needs correction, we can choose x = 0and y = 0for simplicity

    Determine the adjustment vector e: We want d′ = d + e such that the z- component of d′ is zero: 6 +z = 0 =⇒ z = −6 Thus, e = (x, y,−6). Since only the vertical displacement needs correction, we can choose x = 0and y = 0for simplicity. Therefore: e = (0, 0, −6) The magnitude of...

  15. [25]

    Optimal Lineup to Maximize Average ER: • Sort the 10 players by their Efficiency Rating (ER). • Select the top 5 players with the highest ER values: ER1 ≥ ER2 ≥ ER3 ≥ ER4 ≥ ER5 • Calculate the maximum average ER: ER1 + ER2 + ER3 + ER4 + ER5 5 • According to the problem, the ma...

  16. [26]

    - If p = 1, the team is certain to win each game

    Projecting the Team’s Winning Probabil- ity: Using the regression model: P (win) = 0.5 + 0.1 × 25 = 0.5 + 2.5 = 3.0 Since probabilities cannot exceed 1, the model saturates at: P (win) = 1 Probability of Winning a Best-of-7 Series: - The team must win 4 out of 7 games. - If p ...

    2019

  17. [2024]

    Nuo Chen, Zinan Zheng, Ning Wu, Ming Gong, Dong- mei Zhang, and Jia Li

    xcot: Cross-lingual instruction tuning for cross-lingual chain-of-thought reasoning. Nuo Chen, Zinan Zheng, Ning Wu, Ming Gong, Dong- mei Zhang, and Jia Li. 2024. Breaking language barriers in multilingual mathematical reasoning: In- sights and observations. In Findings of the...

    2024 arXiv

  18. [2025]

    In International Conference on Learning Representations

    Allam: Large language models for arabic and english. In International Conference on Learning Representations. Abdelaziz Bounhar and Abdeljalil El Majjodi. 2025. Al-atlas: A causal language model for moroc- can darija. https://huggingface.co/atlasia/ Al-Atlas-0.5B/. Linzheng Ch...

    2025

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.