BLOOM: A 176B-Parameter Open-Access Multilingual Language Model

1. Introduction

• Pretrained language models often improve performance with less labeled data, can perform useful tasks without further training, and tend to benefit from increased scale.The passage describes scaling gains as sometimes predictable and sometimes sudden.Source: §1 · pp. 3–4
• Training costs and limited public releases have excluded much of the research community, while most large language models have primarily used English-language text.The passage also notes environmental concerns about large-model training and cites Chinese- and Korean-language exceptions.Source: §1 · p. 4
• BLOOM is a 176-billion-parameter model trained on 46 natural and 13 programming languages, developed by hundreds of researchers.The paper describes its design across the training dataset, architecture and objective, and distributed-learning engineering, as well as an analysis of its capabilities.Source: §1 · p. 4

2.1 Language Modeling

• Autoregressive language modeling predicts each next token from preceding tokens; Transformers later outperformed recurrent networks and became the de facto language-model architecture.Earlier n-gram models grew exponentially with sequence length and could not assign probabilities to unseen token sequences.Source: §2.1 · pp. 4–5
• Transfer learning pretrains a model before downstream finetuning, while few- and zero-shot approaches use pretrained models without subsequent task-specific training, often through prompts and demonstrations.Prompts provide natural-language task descriptions and can include examples of input-output behavior.Source: §2.1 · p. 5
• Concentrated resources, high computational and carbon costs, and limited stakeholder input raise concerns about access, emissions, governance, and whose values guide LLM development.The text also warns that industry-driven narratives can inflate expectations and misdirect research and policy priorities.Source: §2.1 · pp. 5–6

2.2 BigScience

• BigScience coordinated BLOOM as an open research collaboration: over 1,200 people registered, hundreds contributed to released artifacts, and 30 working groups organized project research.Participants came from varied technical and social-science backgrounds, and 38 countries were represented.Source: §2.2 · p. 6, Figure 1 · p. 7
• BigScience used a collaboratively designed Ethical Charter and research on regulations beyond the US to guide the project, emphasizing inclusivity, diversity, openness, reproducibility, and responsibility.These values were reflected in dataset curation, modeling, engineering, evaluation, and social-impact work.Source: §2.2 · p. 7

3.1 Training Dataset

• ROOTS combines 498 datasets into 1.61 terabytes spanning 46 natural and 13 programming languages, with language inclusion and source selection guided by community expertise.The corpus also included OSCAR, which constituted 38% of its contents.Source: §3.1 · p. 7, §3.1.2 · pp. 10–11

3.1.3 Data Preprocessing

• Curation aimed to counter opaque, heuristic filtering by prioritizing human involvement and local and language expertise in data selection and documentation.Quality indicators targeted text written by humans for humans, with language-specific thresholds and manually selected indicators; the team also removed near-duplicates and redacted identifiable information in OSCAR.Source: §3.1 · p. 9, §3.1.3 · pp. 12–13

3.1.1 Data Governance

• BigScience pursued data governance that prioritized rights-holder agency, but used practical measures including seeking provider permission, keeping sources traceable, and releasing sources compositely.The project could distribute 223 of ROOTS’s 498 components, subject to licensing, privacy, and custodian agreements.Source: §3.1.1 · pp. 9–10

3.1.4 Prompted Datasets

• BLOOMZ was trained on xP3, a prompted-task collection covering 83 datasets, 46 languages, and 16 tasks; xP3mt additionally provided machine-translated prompts.The xP3 language distribution mirrors ROOTS (Figure 4).Source: Figure 4 · p. 13, §3.1.4 · p. 14

3.2.1 Design Methodology

• Architecture choices were ablated at smaller scales and evaluated for zero-shot generalization, while mixture-of-experts and state-space architectures remained out of scope.The team favored model families with scaling evidence and public tooling; it cited limited suitable GPU codebases for mixture-of-experts and underperformance on natural-language tasks at the time for state-space models.Source: §3.2.1 · pp. 14–15

3.2.2 Architecture and Pretraining Objective

• Causal decoder-only models performed best immediately after pretraining in the architecture and objective evaluation, supporting BLOOM’s selection of that design.The study also found these models could be adapted more efficiently after pretraining to a non-causal architecture and objective.Source: §3.2.2 · p. 15

3.2.3 Modeling Details

• BLOOM adopted ALiBi, which improved training smoothness and downstream performance over learned and rotary embeddings, and embedding-layer normalization to mitigate training instability.The layer normalization penalized zero-shot generalization in preliminary experiments; those experiments used float16, whereas final training used bfloat16.Source: §3.2.3 · p. 16

3.3 Tokenization

• Tokenizer design targeted no more than 10 percentage points of fertility degradation versus monolingual tokenizers; the final vocabulary contained 250,680 items.After duplicate lines were removed from tokenizer training data, the team used language sampling ratios matching the training data; 200 vocabulary items were reserved for possible future uses.Source: §3.3 · pp. 17–18
• BLOOM’s byte-level BPE starts from bytes to avoid unknown tokens and maximize vocabulary sharing, while text is left unnormalized to preserve generality.Pre-tokenization preserves spaces and line breaks, avoids English-centric splits and does not split numbers, which caused issues in Arabic and code; Unicode normalization reduced fertility by no more than 0.8% across considered languages.Source: §3.3 · p. 18

3.4.1 Hardware

• 3.5 months and 1,082,990 compute hours were used to train BLOOM on 384 A100 GPUs across 48 nodes, with four spare nodes maintained for hardware failures.The system used Jean Zay, with intra-node NVLink and inter-node Omni-Path links.Source: §3.4.1 · p. 18

3.4.2 Framework; 3.4.4 Fused CUDA Kernels

• 156 TFLOPs in the fastest A100 configuration reached half the 312-TFLOPs theoretical peak using combined data, tensor, pipeline, and ZeRO stage 1 parallelism.Megatron-DeepSpeed combined the approaches to scale training across hundreds of GPUs. Custom fused CUDA kernels optimized LayerNorm, scaling, masking, softmax, and GeLU operations to reduce memory-transfer overhead.Source: §3.4.2 · pp. 19–20, §3.4.4 · p. 20

3.4.3 Floating Point Format

• Bfloat16 mixed-precision training resolved the instability problem seen in earlier 104B-parameter experiments using IEEE float16 on V100 GPUs.Gradient accumulation and softmax were computed in float32, while other operations used lower precision.Source: §3.4.3 · p. 20

3.4.5 Additional Challenges

• Hardware failures averaged 1–2 GPUs per week, but spare nodes and checkpoints every three hours kept throughput from being significantly affected.Data-loader deadlocks and disk-space issues caused 5–10-hour downtimes; the team reported one loss spike, from which the model recovered swiftly.Source: §3.4.5 · pp. 20–21

3.5 Training; Multitask Finetuning

• BLOOM training used a cosine learning-rate decay over 410B tokens with 375M warmup tokens, while large models received 25B additional tokens on repeated data.ROOTS contained around 341B tokens; because warmup plus decay tokens exceeded the available training tokens, the learning-rate decay endpoint was never reached. BLOOMZ retained BLOOM architecture hyperparameters, was finetuned for 13B tokens, and its best checkpoint was selected on a separate validation set; contrastive finetuning produced embedding models for multilingual retrieval and semantic similarity.Source: §3.5 · p. 22, Multitask Finetuning · p. 22

3.5.1 Carbon Footprint

• 81 tons of CO2eq were attributed to BLOOM training, with 25 tons from training energy and 45 tons from idle equipment and cluster consumption.BLOOM used 433 MWh versus OPT’s 324 MWh, yet its training emissions were 25 tons versus OPT’s 70 tons, alongside grid intensities of 57 versus 231 gCO2eq/kWh. Final training accounted for approximately 37% of total project-computation emissions; API deployment on a 16-GPU GCP instance was estimated at 20 kg CO2eq per day, varying by use case.Source: §3.5.1 · p. 23, Table 4 · p. 23

3.6 Release

• BLOOM was released with a Model Card and a no-charge Responsible AI License that separates code and model licensing and imposes 13 behavioral-use restrictions.The source code is under Apache 2.0; the model license permits use subject to its terms and explicitly defines downstream use including prompting, finetuning, distillation, logits, and probability distributions.Source: §3.6 · p. 24

4. Evaluation

• The broader evaluation targeted zero- and few-shot performance across SuperGLUE, translation, and summarization, as well as code generation, text embeddings, multilingual probing, and multitask finetuning.The authors selected prompt-based settings because prompt adaptation and in-context learning were more common than finetuning for models at this scale.Source: §4 · p. 24

4.1.1 Prompts

• Prompts were created before BLOOM’s release without model-based refinement to approximate reproducible, realistic zero- or one-shot use rather than best-case prompt tuning.The authors argue that trial-and-error prompt refinement is harder to reproduce and less representative of ordinary use or true zero-shot learning.Source: §4.1.1 · p. 25
• The evaluation suite generated multiple prompts per task with promptsource (Bach et al., 2022), following crowdsourced prompt generation by Sanh et al. (2022), with peer review for clarity and consistency.Prompt length and style varied substantially.Source: §4.1.1 · p. 25

4.1.2 Infrastructure

• The Prompted Language Model Evaluation Harness extends EleutherAI’s harness with promptsource (Bach et al., 2022) and was released as an open-source library.The framework ran experiments and aggregated results.Source: §4.1.2 · p. 25

SuperGLUE

• The SuperGLUE evaluation used seven English-only tasks, five randomly sampled prompts per task, and maximum log likelihood over candidate label strings for predictions.The remaining tasks were excluded because they required an order of magnitude more compute than the selected tasks combined; the subset supports comparison with prior English-only work.Source: SuperGLUE · p. 26
• For SuperGLUE, T0 (Sanh et al., 2022) was noted as an exception in reporting zero- and one-shot results, but its instruction tuning made it not directly comparable to BLOOM or OPT.The selected tasks were English-only and included to facilitate comparison with prior work, which had mainly focused on English-only models.Source: SuperGLUE · p. 26
• Translation evaluation covered WMT14 en↔fr and en↔hi, Flores-101, and DiaBLa, using sacrebleu BLEU with dataset-specific tokenization and greedy decoding.Generation stopped at EOS or, for one-shot cases, additionally at the specified separator; maximum lengths were 64 tokens for WMT14 and 512 for Flores-101 and DiaBLa.Source: SuperGLUE · p. 26
• WikiLingua summarization tested source-language generation in nine languages, including Arabic, English, Spanish, French, Hindi, Indonesian, Portuguese, Vietnamese, and Chinese.This design examined BLOOM’s multilingual summarization ability rather than limiting evaluation to English output.Source: Summarization · p. 26

Summarization

• Summarization is evaluated with ROUGE-2, ROUGE-L, and Levenshtein distance, using a Flores-101 SentencePiece tokenizer, 3,000 uniformly sampled GEM test examples, and the MT decoding procedure.The authors note minimal performance differences between the subsamples and full test sets.Source: Summarization · pp. 26–27

4.1.4 Baseline Models

• Where language support permits, GPT-style baselines include mGPT (Shliazhko et al., 2022), GPT-Neo (Black et al.), GPT-J-6B, GPT-NeoX (Black et al., 2022), OPT (Zhang et al., 2022), and XGLM (Lin et al., 2021).mGPT was trained on 60 languages; XGLM is multilingual, and the GPT-Neo family and OPT are trained on dataset mixtures described in the section.Source: §4.1.4 · p. 27
• Other baselines are T0 (Sanh et al., 2022) and mTk-Instruct (Wang et al., 2022b), T5 variants that underwent multitask prompted finetuning.T0 used datasets from P3; mTk-Instruct used Super-NaturalInstructions.Source: §4.1.4 · p. 27
• The baseline set also includes M2M (Fan et al., 2021), AlexaTM (Soltan et al., 2022), Codex (Chen et al., 2021), and GPT-fr.M2M translates between 100 languages; AlexaTM is an encoder-decoder model, Codex is finetuned on GitHub code, and GPT-fr was trained on French text.Source: §4.1.4 · p. 27

4.2 SuperGLUE

• BLOOM gains more than comparable models from zero-shot to one-shot SuperGLUE prompting, generally matching or surpassing OPT despite often trailing it zero-shot.Performance increases are slight and inconsistent overall, while variability across prompts and models falls in the one-shot setting.Source: §4.2 · p. 28, Figure 7 · p. 28
• On SuperGLUE, BLOOM and the compared models perform well above chance on BoolQ and CB, while average performance across prompts on other tasks hovers around chance.T0 is an exception with strong performance, but its multitask fine-tuning makes it not directly comparable to the other models.Source: §4.2 · pp. 27–28, Figure 7 · p. 28
• Across model scales, BLOOM and OPT improve only very slightly, with signal appearing only above 2 billion parameters and no consistent family-wide difference.In one-shot tests, BLOOM-176B leads OPT-175B on Ax-b, CB, WSC, and WiC, and matches it on the other tasks.Source: §4.2 · p. 28, Figure 8 · p. 29

4.3 Machine Translation

• The machine-translation evaluations cover WMT zero- and one-shot tests, DiaBLa dialogue-context tests, and one-shot Flores-101 translation across resource levels and language directions.The Flores tests use the “xglm-source+target” prompt with a random dev-set example; language pairs are grouped by resource level, including related Romance pairs.Source: §4.3.1 · p. 29, §4.3.2 · p. 30, §4.3.3 · p. 31

4.3.1 WMT

• M2M-100 scored 43.8 BLEU for English→French and 40.4 for French→English, versus BLOOM’s 34.2 and 35.4 in one-shot WMT translation.The more verbose prompts tended to perform best; “version-target” was consistently better, while “gpt3-target” and “xglm-source+target” performed poorly, especially zero-shot. Over-generation and incorrect output language were prominent zero-shot problems, both greatly improving with more few-shot examples.Source: §4.3.1 · pp. 29–30, Table 6 · p. 30

4.3.2 DiaBLa

• DiaBLa’s automatic scores are inconclusive: previous-dialogue context raises BLEU relative to a random example but lowers COMET, although predictions show evidence of context-sensitive translation choices.The comparison used the “xglm-source+target” prompt in a one-shot setting and evaluated both original outputs and outputs truncated to address over-generation.Source: §4.3.2 · p. 30

4.3.3 Flores

• In 1-shot translation, BLOOM’s results were often comparable to and sometimes better than supervised 615M-parameter M2M-100, with good results across high-resource and high-to-mid-resource pairs, even across scripts (Fan et al., 2021).Results were also comparable in many cases to 32-shot AlexaTM results for high-resource language pairs.Source: §4.3.3 · p. 32
• Flores-101 results are poor for Swahili–Yoruba, whose languages each had fewer than 50k training tokens, but good across Romance pairs, including unseen Galician.The contrast raises a concern about quality for underrepresented low-resource languages that were included in training.Source: §4.3.3 · p. 32

• BLOOM outperforms OPT-175B on one-shot multilingual summarization, and performance increases with model parameter count (Figure 9).Source: §4.4 · p. 32
• The authors suspect BLOOM’s multilingual-focused training contributes to its stronger multilingual summarization performance.Source: §4.4 · p. 32
• ROUGE-2 is reported for comparability with prior work because alternatives for generation evaluation are lacking, but it can understate summary quality.Source: §4.4 · p. 32

4.5 Code Generation

• On HumanEval, pretrained BLOOM models perform similarly to similarly sized GPT models trained on the Pile.ROOTS contains around 11% code, compared with around 13% in the Pile; both corpora have similar code sources and proportions.Source: §4.5 · p. 32
• Codex models, which were solely finetuned on code, are significantly stronger than other models.Source: §4.5 · p. 32
• Multitask-finetuned BLOOMZ models do not significantly improve over BLOOM, possibly because xP3 lacks substantial pure code-completion data.Instead, xP3 includes code-related tasks such as estimating the time complexity of a Python code snippet.Source: §4.5 · p. 32

4.6 HELM benchmark

• On HELM’s 5-shot English-only tasks, BLOOM is roughly on par with GPT3-davinci v1 and J1-Grande v1, but behind newer monolingual models.The newer models named are InstructGPT davinci v2, Turing NLG v2, Anthropic-LM v4-s3, and OPT.Source: §4.6 · p. 33
• On HELM, BLOOM is not very well calibrated but is quite robust.Source: §4.6 · p. 33
• On HELM, BLOOM is among the best models for fairness, slightly more toxic than average in English, and average for bias.Source: §4.6 · p. 33

4.7 Multitask Finetuning

• 33% for NLI (XNLI) and 50% for coreference resolution (XWinograd) are the random baselines that pretrained BLOOM and XGLM perform near.Source: §4.7 · p. 33
• Multilingual multitask finetuning with xP3 significantly improves BLOOMZ zero-shot performance on held-out NLI, coreference, and sentence-completion tasks.The comparison includes pretrained BLOOM and XGLM and multitask-finetuned BLOOMZ, T0, and mTk-Instruct.Source: §4.7 · pp. 33–34
• T0 performs badly on the depicted multilingual datasets because it is a monolingual English model.Source: §4.7 · p. 34
• xP3-finetuned models also outperform T0 on English datasets when controlling for size and architecture, likely because xP3 has more diverse datasets and prompts than P3 (Muennighoff et al., 2022b; Chung et al., 2022).Source: §4.7 · p. 34

4.8 Embeddings

• On two multilingual MTEB datasets, SGPT-BLOOM-7.1B-msmarco[36] achieves state-of-the-art performance on several classification and semantic textual similarity splits.The models use contrastive finetuning.Source: §4.8 · pp. 34–35
• SGPT-BLOOM-7.1B-msmarco[36] is an order of magnitude larger than the displayed multilingual MiniLM[37] and MPNet[38] models.Source: §4.8 · p. 35
• SGPT-BLOOM-1.7B-nli[39] performs significantly worse, likely because it has fewer parameters and shorter finetuning on a much smaller dataset.The passage identifies NLI as much smaller than MS-MARCO.Source: §4.8 · p. 35
• ST5-XL[40], the largest non-BLOOM model at 1.2 billion parameters, performs poorly on non-English data because it is English-only.The displayed languages are part of BLOOM's pretraining corpus; more languages and datasets are available on the MTEB leaderboard[41].Source: §4.8 · pp. 35–36

4.9.1 Method

• The study probes 17 languages from 7 families and 38 morphosyntactic features using layer-pooled sentence embeddings and logistic-regression classifiers.It uses the 1.7B and 176B BLOOM models, evaluates weighted F1 across three random-seed runs, and compares against random guessing and TF-IDF n-gram baselines.Source: §4.9.1 · pp. 37–38

4.9.2 Results

• BLOOM-1B7 performs on par with or better than BLOOM, and both models outperform count-based baselines across the probing results.The reported averages cover probing tasks and experiment runs within each language.Source: Table 12 · p. 39, §4.9.2 · p. 39
• Both models infer Mood and Person well across languages, Number, NumType, and Voice moderately, and other categories less well; Bengali, Wolof, and Yoruba score lowest.The authors suggest that categories with more language-dependent values, such as Case, may be harder to infer than categories with more shared values.Source: Figure 12 · p. 39, §4.9.2 · p. 40
• BLOOM-1B7 results correlate strongly with language family and probing and pretraining dataset sizes, while BLOOM-1B7 scores significantly higher than BLOOM (p < 0.01).BLOOM is reported to perform more stably across languages regardless of pretraining data amount, possibly indicating better generalization with more parameters.Source: §4.9.2 · p. 40, Table 13 · p. 40
• Future work includes probing languages outside pretraining, studying under-resourced languages, and examining layer-specific representations and how linguistic properties emerge during pretraining.The authors also caution that BLOOM-1B7’s stronger average results and stronger dataset-size correlations may imply weaker generalization to under-resourced languages than the larger model.Source: §4.9.2 · pp. 40–41

• The evaluation adapted multilingual CrowS-Pairs—revised English pairs and a French version—to autoregressive models, using prompts that ask which minimal-pair statement seems likely and realistic.The dataset assesses systematic preference for stereotyped over non-stereotyped statements.Source: §4.10 · p. 41
• CrowS-Pairs accuracy was homogeneous across bias categories, unlike prior masked-language-model findings, yet differed significantly from 50 overall in both languages and in several categories.The category results are averaged over eight runs; asterisks indicate significance at p < .05.Source: §4.10 · pp. 41–42, Table 14 · p. 42
• BLOOM’s overall prompt accuracy was close to .50 in English and French, suggesting limited overall bias and similar behavior across the two languages.Figure 13 also shows results for the two smallest BLOOM models and monolingual GPT models of comparable size.Source: §4.10 · p. 41, Figure 13 · p. 42
• The CrowS-Pairs version addresses some validity issues raised by Blodgett et al. (2021), adds 200 stereotype pairs based on French speakers, and had results not significantly different from the original in Névéol et al. (2022).Source: §4.10 · p. 42
• CrowS-Pairs’ original validation does not naturally apply to this evaluation, making comparison with other results difficult; stronger assessment requires other bias measures and coverage of all model languages.Very little multilingual bias-assessment material is available.Source: §4.10 · p. 43
• The evaluation covers only situations, languages, and variants in multilingual CrowS-Pairs; curated training resources may unevenly capture language variation and thereby affect equitable representation and legitimize some variants over others.The authors distinguish these findings from broader model use and point to Raji et al. (2021) for further exploration.Source: §4.10 · p. 43

5. Conclusion

• BLOOM is a 176B-parameter open-access multilingual model whose competitive performance improves after multitask finetuning.BigScience, a collaboration of hundreds of researchers, trained it on the French government-funded Jean Zay supercomputer for 3.5 months.Source: §5 · p. 43
• The authors hope BLOOM’s release will unlock applications and research directions, while their account of BigScience will help researchers organize collaborative projects and broaden participation.They emphasize that such projects can achieve results beyond an individual research group and enable people with different backgrounds to share ideas.Source: §5 · p. 43

Major Contributors

• Major Contributors lists people essential to BLOOM or those who spent more than 20% of their time on the BigScience effort.The category’s author order is randomly shuffled, except that Teven Le Scao is intentionally listed first.Source: §6 · p. 43

Dataset and Tokenization

• Dataset contributors handled data sourcing, organization, and processing, while Tokenization contributors built the BLOOM tokenizer.The Dataset category includes authors of Lauren¸con et al. (2022), McMillan-Major et al. (2022), and Jernite et al. (2022); Tokenization includes authors of Mielke.Source: §6 · p. 44

Prompt Engineering and Architecture and Objective

• Prompt Engineering contributors wrote, edited, and reviewed dataset prompt templates, while Architecture and Objective contributors ran experiments to determine BLOOM’s architecture and training objective.The categories include authors of Sanh et al. (2022), Bach et al. (2022), Muennighoff et al. (2022b), Wang et al. (2022a), and Le Scao et al. (2022).Source: §6 · p. 44

Engineering and Evaluation and interpretability

• Engineering contributors supported BLOOM’s code and training infrastructure on Jean Zay, while Evaluation and interpretability contributors helped evaluate the model.The Evaluation and interpretability category also includes authors of Talat et al. (2022).Source: §6 · p. 44

Broader Impacts, Applications, and Organization

• Broader Impacts covers the ethical charter, license, model card, privacy, social impacts, and carbon footprint; Applications covers application-focused working groups; Organization covers BigScience coordination.Applications includes authors of Fries et al. (2022b), Fries et al. (2022a), and De Toni et al. (2022); Organization includes authors of Akiki.Source: §6 · p. 44

Appendix A. Prompts

• The appendix supplies evaluation prompts as raw templates and applied examples, using double curly brackets for sample content.It also points readers to PromptSource (Bach et al., 2022).Source: Appendix A · p. 65

A.1.1 Data example; A.2.1 Data example; GPT-3-prompt-with-label

• Examples test pronoun reference and word-sense judgments, with a separate template asking whether a labeled word has the same sense in two sentences.Source: §A.1.1 · p. 66, §A.2.1 · p. 66, GPT-3-prompt-with-label · p. 67

A.3.1 Data example; A.4.1 Data example; A.5.1 Data example

• Other examples pose a definition-based question, test whether a statement follows from a passage, and ask whether a speaker has spoken to someone again.Source: §A.3.1 · p. 67, §A.4.1 · p. 68, §A.5.1 · p. 68

A.6.1 Data example; A.7.1 Data example

• A pronoun-resolution example selects “The city councilmen” as the sentence completion, while another example chooses the cause “It was fragile” for bubble-wrapped packaging.Source: §A.6.1 · p. 69, §A.7.1 · p. 70

A.7.2 Prompts

• A.7’s reusable causal prompt template offers two answer choices and inserts “because” for cause questions or “so” otherwise.Source: §A.7.2 · p. 70

A.8.1 Data example; A.9 WMT; A.9.1 Data example; a_good_translation-l1-l2-source+target

• The appendix notes that XStoryCloze and Story Cloze samples are not public, and its WMT prompts cover zero-shot and one-shot English–French and English–Hindi directions with language-specific variants.The WMT example presents a source sentence followed by a requested target-language translation.Source: §A.8.1 · p. 71, §A.9 · p. 71, §A.9.1 · p. 71

A.10 DiaBLa; A.10.1 Data example; xglm-source+target

• DiaBLa examples use English–French source-target prompts, and few-shot examples may follow either direction or be controlled to match or oppose the current direction.Source: §A.10 · p. 72, §A.10.1 · p. 72, xglm-source+target · p. 72

A.11 Flores-101 (MT); A.11.1 Data example; xglm-French-Catalan-source+target; A.12.1 Data example; demontraste_or_violate

• The remaining examples show pair-specific French-to-Catalan translation prompts and a choice between a stereotype demonstration and a stereotype violation.Source: §A.11 · p. 72, §A.11.1 · p. 73, xglm-French-Catalan-source+target · p. 73