Pith. sign in

REVIEW 6 major objections 6 minor 30 references

SPOT: Bridging Natural Language and Geospatial Search for Investigative Journalists

T0 review · 6 major / 6 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 SPOT is the first system to reliably translate plain-language scene descriptions into accurate OpenStreetMap queries, using a fine-tuned LLaMA 3 parser and a semantic tag-bundling index.

desk verdict Solid open-source parser for natural-language OSM queries, but the 'reliable access' headline outruns the evidence—no end-to-end evaluation and ~50% relation accuracy. read the letter →

arxiv 2506.13188 v1 pith:JB4ZWWKC submitted 2025-06-16 cs.IR cs.CLcs.HC

Lynn Khellaf , Ipek Baris Schlicht , Tilman Mirass , Julia Bayer , Tilman Wagner , Ruben Bouwmeester This is my paper
classification cs.IRcs.CLcs.HC
keywords OpenStreetMapnaturallanguageinterfacegeospatialsearchLLMfine-tuninginvestigativejournalismsemanticparsingYAMLquerygenerationgeolocationverification
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

SPOT is an open-source natural language interface that lets non-technical users, especially investigative journalists, search OpenStreetMap by describing what they see in plain text instead of writing OverpassQL queries. The paper's central claim is that a compact open-weight language model fine-tuned on synthetic scene descriptions can translate such text into a structured YAML scene description—area, entities, properties, and spatial relations—with higher accuracy on entities, properties, and relations than zero-shot or one-shot GPT-4o and a fine-tuned mT5 baseline, and with fewer hallucinated or omitted objects. A semantic bundling step then maps the generic names in the YAML to concrete OSM tags, so model output stays independent of OSM tagging changes. If the claim holds, SPOT would be the first reliable way for journalists without query-language expertise to geolocate from scene descriptions, lowering the technical barrier to verification work. The authors present results on a 195-query benchmark built from real journalist queries, with LLaMA 3 selected as the production parser because it yields consistently parseable output.

What carries the argument

The load-bearing mechanism is the two-stage decoupling of language understanding from OSM-specific knowledge. The first stage is a fine-tuned LLaMA 3 model, trained on synthetic sentence–YAML pairs generated by GPT-4o from random scene configurations, which outputs a tag-free YAML scene description with area, entities, properties, and distance or contains relations. The second stage is semantic bundling: a static list of bundles that groups visually similar OSM tags under natural-language descriptors, indexed with BM25 and SBERT embeddings so that typos and paraphrases retrieve the right bundle; the backend then replaces the generic YAML names with the bundle's OSM tags before querying the geodatabase. This decoupling means OSM tag changes can be absorbed by updating bundles without retraining the model, and vague spatial terms are converted to fixed distance values during data generation, making relation extraction a concrete prediction task rather than an open-ended interpretation.

What would settle it

Ask a group of investigative journalists to mark on a map what area they mean by 'near' and 'nearby' in actual geolocation prompts, and compare those marked areas with the fixed 100 m and 2000 m values used in training; if the intended distances scatter widely around those values, then the spatial-relation layer will place correct entities at wrong distances. A complementary check is to run SPOT end-to-end on queries with known ground-truth locations and count how often the returned pin matches the true location.

Watch

Extended reading notes

Core claim

The paper sets out to show that natural language access to OpenStreetMap can be made reliable enough for investigative work by separating the problem into two layers. In the first layer, a fine-tuned LLaMA 3 model converts free-form user text into a YAML intermediate representation containing the search area, the entities involved, their properties, and the spatial relations between them; this is a scene description rather than an OSM query. In the second layer, entity and property names are passed through a semantic search over a hand-built list of OSM tag bundles, so that colloquial, misspelled, or ambiguous terms are resolved to the best-matching OSM tags before the query is executed against a geospatial database. On a 195-sample benchmark built from queries produced by journalists and fact-checkers, the fine-tuned open models outperform GPT-4o and mT5 on entity, property, and relation extraction, and LLaMA 3 is the only fine-tuned model that consistently emits parseable YAML; the paper reports that this makes SPOT the first system to achieve reliable natural-language access to OSM data at this accuracy level.

Load-bearing premise

The load-bearing assumption is that the fixed translation table for vague spatial words—'next to' as 50 m, 'near' as 100 m, 'nearby' as 2000 m—matches what a user actually means when speaking; if journalists mean different distances, the parsed query will search the wrong places even when every entity and area is correct.

Editorial extensions

If this is right

  • A journalist who types a scene description such as 'find an Italian restaurant next to a Carrefour in Paris' can receive map results without knowing OverpassQL or OSM tag conventions.
  • Fine-tuned open-weight models can beat a much larger closed model on this structured extraction task, so the interface does not depend on a proprietary API.
  • Because the LLM output is tag-free and bundle search resolves tags separately, OSM tagging inconsistencies and future tag changes can be handled by editing the bundle list rather than retraining the model.
  • If the benchmark reflects real use, SPOT's lower hallucination and omission counts mean fewer misleading locations are shown to fact-checkers under time pressure.

Reading between the lines

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

  • The same tag-free YAML spine could be retargeted to geodata sources other than OSM or to different query back ends by replacing the bundle index, without retraining the language model; this is an extension the paper does not test.
  • The fixed distance table for vague words ('near' as 100 m, 'nearby' as 2000 m) is a user-intent assumption; a natural extension would calibrate or personalize these values per region or per journalist.
  • The 195-query benchmark measures parsing accuracy, not end-to-end locational success; a field test with known ground-truth locations would show whether correct YAML translates into correct pins on the map.
  • The synthetic data pipeline could serve as a template for building robust natural-language interfaces to other structured, tag-heavy datasets, since it injects typos, grammar errors, personas, and non-Latin scripts at scale.
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

6 major / 6 minor

Summary. The paper describes SPOT, an open-source natural language interface to OpenStreetMap designed for investigative journalists. SPOT uses a fine-tuned LLaMA 3 model to convert unstructured user queries into a YAML-based scene representation with areas, entities, properties, and spatial relations; this YAML is then mapped to OSM tag bundles via hybrid BM25/SBERT retrieval and executed against a PostGIS database. The authors introduce a synthetic data generation pipeline that produces training pairs by asking GPT-4o to verbalize random YAML configurations, and they evaluate several open-weight LLMs plus GPT-4o and mT5 on a benchmark of 195 real journalist queries, reporting per-field accuracy (area, entity, property, relation) and hallucination counts. They claim, to their knowledge, the first system to achieve reliable natural language access to OSM data at this level of accuracy, with state-of-the-art query interpretation performance.

Significance. The paper's strengths include full open-source release (code and model weights), a real user benchmark collected from journalists and fact-checkers, and a systematic comparison of multiple open-weight LLMs. If validated end-to-end, SPOT would be a useful tool for lowering the technical barrier to OSM querying. However, the claimed 'reliable natural language access' is not established by the evidence: the evaluation measures only intermediate YAML parsing accuracy, not whether the final map results match user intent, and the relation and property accuracies are around 50%. The arbitrary mapping of vague spatial terms to fixed distances is used as ground truth without a user study, and the benchmark was filtered by ill-defined criteria. These gaps prevent the paper from supporting its headline claims, but they are addressable with additional evaluation and more modest framing.

major comments (6)
  1. [Section 4.2 and Conclusion] The headline claim of 'reliable natural language access to OSM data' is supported only by per-field YAML matching scores in Table 6, not by any end-to-end retrieval evaluation; the Conclusion itself defers 'comprehensive end-to-end evaluations' to future work, and the deployed model scores 48.05% on relations and 50.00% on properties. Because the actual user-facing outcome depends on bundle matching, PostGIS execution, and display, the intermediate parsing accuracy is a proxy, not a measure of whether a journalist finds the intended place. Please either add an end-to-end evaluation on the benchmark (e.g., whether the top-k returned locations match the ground-truth coordinates) or temper the abstract and conclusion claims to the parser-level evidence.
  2. [Appendix A.2, Table 8] The mapping of vague spatial terms to concrete distances (e.g., 'next to' = 50 m, 'near' = 100 m, 'nearby' = 2000 m) is used as ground truth in both the synthetic training data and the benchmark evaluation, but no user study or evidence shows these values correspond to what journalists mean. If a journalist intending 'near' as 10 m receives results within 100 m, the parser may score correct per the metric while the system underperforms. Please provide a user study or at least a sensitivity analysis showing the reported relation accuracies are stable under alternative plausible mappings.
  3. [Section 4.1, Benchmarking Dataset] The benchmark was filtered 'based on how well the queries aligned with the OSM database structure and its resulting limitations,' but the filter criteria, the number of excluded queries, and examples of exclusions are not provided. This selection can strongly bias the reported accuracies upward, making them non-representative of the real query distribution. Please report the full pre-filter corpus, the filtering rules, and the effect of filtering on the metric, or justify that the filtered set is the appropriate target distribution.
  4. [Section 4.2, Results] Table 6 reports point estimates without confidence intervals or significance tests; on 195 samples with multiple categories, the 4-5 point differences between models (e.g., LLaMA 3 at 48.05% vs. Qwen2.5 at 52.60% relations) are likely within sampling noise. The hypothesis that 'Qwen2.5 having the most parameters could indicate that relation identification is a task that requires advanced reasoning skills' is therefore speculative. Please report confidence intervals and significance tests (or a bootstrap analysis), and avoid causal interpretations of model size without controlling for other factors.
  5. [Section 4.2, Parsability] The deployed parser is LLaMA 3 because 'only LLaMA 3 and GPT-4o consistently produced parsable output,' yet the parsability rates are not reported, and LLaMA 3 is not the best model on the main evaluation metrics (Phi leads on properties at 53.90%, Qwen2.5 on relations at 52.60%). Without parsability numbers, it is unclear whether the deployed configuration is actually the best system, and the claim of 'state-of-the-art performance in query interpretation with fine-tuned LLaMA 3' is stronger than the evidence. Please report parsability rates for all models and explain the selection trade-offs.
  6. [Section 2.2 and Section 4.2] The paper claims state-of-the-art performance but does not compare against existing OSM-specific natural language interfaces such as NLMaps (Lawrence and Riezler, 2016) and OverpassT5 (Staniek et al., 2024), which are cited as prior work. Since those systems also interpret natural language for OSM, a direct comparison (or a clear argument why the tasks differ in a way that makes comparison impossible) is necessary to substantiate the 'state-of-the-art' claim. At minimum, the discussion should situate SPOT's per-field accuracies against the task formulations of these prior systems.
minor comments (6)
  1. [Throughout] The term 'YAML' is inconsistently rendered as 'YAML' and 'Y AML'; use a single consistent spelling.
  2. [Table 7] The numbers of hallucinated and omitted entities and properties are raw counts; reporting them as rates relative to the total number of ground-truth entities and properties would make the comparison more meaningful.
  3. [Figure 4] The caption should specify exactly what constitutes a 'perfect YAML generation,' since the per-category percentages depend on that definition and the figure is referenced in the analysis.
  4. [References] The reference for LoRA (Hu et al.) is missing a publication year and venue; please add the full ICLR 2022 details.
  5. [Section 4.1, Evaluation Metric] The fixed SBERT cosine similarity threshold of 0.8 for matching entity and property names is arbitrary; a brief justification or sensitivity analysis would clarify the metric's behavior.
  6. [Section 4.2] The statement that the model 'appears to be able to interpret queries in a variety of languages, although this was not further tested' is confusing because Figure 4 includes a 'Non-roman alphabet' category, indicating that non-English queries were in fact tested; please clarify which languages were tested and what 'not further tested' means.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: SPOT's evaluation uses real journalist queries against an explicit YAML schema; the cited self-references are contextual, and the fixed spatial-term distances are design definitions, not fitted predictions.

full rationale

The derivation chain in SPOT is not circular. The core claim—that fine-tuned LLaMA 3 and other open LLMs translate natural-language geospatial queries into structured YAML more accurately than GPT-4o and mT5—is supported by evaluation on a benchmark of 195 real user queries from journalists and fact-checkers that were not used for training. The synthetic training data is generated from the same YAML schema used at inference, but this is standard supervised learning on a declared target representation; it does not make the evaluation measure trivially equal to the training objective, because the test queries are human-produced and were annotated into YAML independently of the generation pipeline. The hand-set vague-spatial-term distances in Appendix A.2 are explicit design definitions (e.g., 'next to' = 50 m) used both in training and in ground-truth annotation; this is a design assumption that may limit external validity, but it is not a fitted parameter disguised as a prediction, since the paper nowhere derives these distances from data or presents them as an empirical result. The only self-citations (Khellaf et al., 2023) are contextual comparisons to the authors' earlier T5 system and a rationale for choosing mT5 as a baseline; they are not load-bearing evidence for the current results. Consequently, no step in the paper reduces by construction to its input, and the central accuracy numbers stand as an internal evaluation of query-interpretation quality.

Assumptions & free parameters 2 free parameters · 4 assumptions · 0 invented entities

The central claim rests on hand-defined distance mappings, an evaluation threshold, and assumptions about OSM coverage and embedding reliability. No new physical or mathematical entities are introduced; the YAML schema and semantic bundles are software artifacts, not postulated entities requiring external evidence.

free parameters (2)
  • Relative spatial term to distance mapping = next to: 50 m; near: 100 m; nearby: 2000 m (Appendix A.2)
    Hand-set distance values for vague terms are used in both synthetic training data and inference. If they do not match user intent, the parsed relations are wrong regardless of entity and area accuracy.
  • SBERT cosine similarity threshold for evaluation matching = 0.8
    Entity and property predictions are counted as correct when their cosine similarity exceeds 0.8. This hand-chosen threshold directly affects the reported accuracy numbers.
assumptions (4)
  • domain assumption OSM tag data has sufficient coverage and consistency for the target geolocation tasks.
    The system relies entirely on OpenStreetMap as its knowledge source. Incomplete or inconsistent OSM data would produce wrong or missing results, a limitation the authors acknowledge.
  • domain assumption SBERT embeddings (all-MiniLM-L6-v2) capture semantic equivalence of entity and property names in the OSM domain.
    Both the evaluation metric and the bundle matcher use this embedding model. If the model is unreliable for OSM-specific terms, both matching and evaluation would be biased.
  • domain assumption GPT-4o-generated synthetic sentences are representative of real journalist queries.
    The training set is synthetic. A user study informed the prompt design, but no quantitative evidence shows the synthetic distribution matches real query patterns, especially for typos and multilingual input.
  • ad hoc to paper The fixed YAML schema captures all information needed to translate a user query into a correct OSM query.
    The system restricts queries to areas, entities, properties, and contains/distance relations. Other query types, such as metadata or complex filtering, are out of scope by construction.

how reviews work

0 comments
Cite this review

Pith. "Pith review of SPOT: Bridging Natural Language and Geospatial Search for Investigative Journalists." pith.science (2026). https://pith.science/paper/JB4ZWWKC

@misc{pith2026250613188,
  author       = {Pith},
  title        = {Pith review of: SPOT: Bridging Natural Language and Geospatial Search for Investigative Journalists},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/JB4ZWWKC}},
  note         = {Machine review of arXiv:2506.13188}
}
read the original abstract

OpenStreetMap (OSM) is a vital resource for investigative journalists doing geolocation verification. However, existing tools to query OSM data such as Overpass Turbo require familiarity with complex query languages, creating barriers for non-technical users. We present SPOT, an open source natural language interface that makes OSM's rich, tag-based geographic data more accessible through intuitive scene descriptions. SPOT interprets user inputs as structured representations of geospatial object configurations using fine-tuned Large Language Models (LLMs), with results being displayed in an interactive map interface. While more general geospatial search tasks are conceivable, SPOT is specifically designed for use in investigative journalism, addressing real-world challenges such as hallucinations in model output, inconsistencies in OSM tagging, and the noisy nature of user input. It combines a novel synthetic data pipeline with a semantic bundling system to enable robust, accurate query generation. To our knowledge, SPOT is the first system to achieve reliable natural language access to OSM data at this level of accuracy. By lowering the technical barrier to geolocation verification, SPOT contributes a practical tool to the broader efforts to support fact-checking and combat disinformation.

Figures

Figures reproduced from arXiv: 2506.13188 by the authors.

Figure 1
Figure 1. Overview of SPOT’s OSM-based pipeline, from tag bundle indexing and semantic search, through artificial [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Sentence length distribution of the generated [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Semantic Diversity Visualization of sentence [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figures from the paper (2 more)
Figure 4
Figure 4. Figure 4: Analysis of LLaMA 3, Mistral, and Phi regarding the ratio of perfect YAML generations various metadata [PITH_FULL_IMAGE:figures/full_fig_p006_4.png]
Figure 5
Figure 5. Figure 5: Zero-shot prompt used to query the LLMs, containing instructions and the YAML layout. The prompt [PITH_FULL_IMAGE:figures/full_fig_p011_5.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

30 extracted references · 11 canonical work pages

  1. [1]

    Marah Abdin, Jyoti Aneja, Hany Awadalla, Ahmed Awadallah, Ammar Ahmad Awan, Nguyen Bach, Amit Bahree, Arash Bakhtiari, Jianmin Bao, Harkirat Behl, et al. 2024. Phi-3 technical report: A highly capable language model locally on your phone. arXiv preprint arXiv:2404.14219

    arXiv 2024

  2. [2]

    Ricardo J. G. B. Campello, Davoud Moulavi, Arthur Zimek, and J\" o rg Sander. 2015. https://doi.org/10.1145/2733381 Hierarchical density estimates for data clustering, visualization, and outlier detection . ACM Trans. Knowl. Discov. Data, 10(1)

    doi:10.1145/2733381 2015

  3. [3]

    Jett Chen. 2025. https://github.com/JettChenT/earthkit Earthkit . Accessed: 2025-02-10

    work page 2025

  4. [4]

    Abhimanyu Dubey, Abhinav Jauhri, Abhinav Pandey, Abhishek Kadian, Ahmad Al - Dahle, Aiesha Letman, Akhil Mathur, Alan Schelten, Amy Yang, Angela Fan, Anirudh Goyal, Anthony Hartshorn, Aobo Yang, Archi Mitra, Archie Sravankumar, Artem Korenev, Arthur Hinsvark, Arun Rao, Aston Zhang, Aur \' e lien Rodriguez, Austen Gregerson, Ava Spataru, Baptiste Rozi \` e...

    arXiv 2024

  5. [5]

    Dawei Gao, Haibin Wang, Yaliang Li, Xiuyu Sun, Yichen Qian, Bolin Ding, and Jingren Zhou. 2024. https://doi.org/10.14778/3641204.3641221 Text-to-sql empowered by large language models: A benchmark evaluation . Proc. VLDB Endow. , 17(5):1132--1145

    arXiv 2024

  6. [6]

    Graylark. 2025. https://geospy.ai/ Geospy . Accessed: 2025-02-10

    work page 2025

  7. [7]

    Daniel Han, Michael Han, and Unsloth team. 2023. Unsloth. http://github.com/unslothai/unsloth

    work page 2023

  8. [8]

    Lora: Low-rank adaptation of large language models

    Edward J Hu, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, Weizhu Chen, et al. Lora: Low-rank adaptation of large language models. In International Conference on Learning Representations

Show all 30 references
  1. [9]

    Joel Jang, Seungone Kim, Seonghyeon Ye, Doyoung Kim, Lajanugen Logeswaran, Moontae Lee, Kyungjae Lee, and Minjoon Seo. 2023. Exploring the benefits of training expert language models over instruction tuning. In International Conference on Machine Learning, pages 14702--14729. PMLR

    2023

  2. [10]

    Albert Q Jiang, Alexandre Sablayrolles, Arthur Mensch, Chris Bamford, Devendra Singh Chaplot, Diego de las Casas, Florian Bressand, Gianna Lengyel, Guillaume Lample, Lucile Saulnier, et al. 2023. Mistral 7b. arXiv preprint arXiv:2310.06825

    2023 arXiv

  3. [11]

    Lynn Khellaf, Ipek Baris Schlicht, Julia Bayer, Ruben Bouwmeester, Tilman Mira , and Tilman Wagner. 2023. Spot: A natural language interface for geospatial searches in osm. Proceedings of OSM Science 2023, page 49

    2023

  4. [12]

    Carolin Lawrence and Stefan Riezler. 2016. Nlmaps: A natural language interface to query openstreetmap. In Proceedings of COLING 2016, the 26th International Conference on Computational Linguistics: System Demonstrations, pages 6--10

    2016

  5. [13]

    Dongjun Lee, Choongwon Park, Jaehyuk Kim, and Heesoo Park. 2025. https://aclanthology.org/2025.coling-main.24/ MCS - SQL : Leveraging multiple prompts and multiple-choice selection for text-to- SQL generation . In Proceedings of the 31st International Conference on Computation...

    2025

  6. [14]

    Leland McInnes, John Healy, Nathaniel Saul, and Lukas Grossberger. 2018. Umap: Uniform manifold approximation and projection. The Journal of Open Source Software, 3(29):861

    2018

  7. [15]

    Bill Meixner. 2025. https://chatgpt.com/g/g-brlHi7t2R-geoguessr-gpt Geoguessr gpt . Accessed: 2025-02-10

    2025

  8. [16]

    OpenAI. 2023. https://arxiv.org/abs/2303.08774 Gpt-4 technical report . Preprint, arXiv:2303.08774

    2023 arXiv

  9. [17]

    OSM contributors . 2017. Planet dump retrieved from https://planet.osm.org. https://www.openstreetmap.org

    2017

  10. [18]

    Qwen, :, An Yang, Baosong Yang, Beichen Zhang, Binyuan Hui, Bo Zheng, Bowen Yu, Chengyuan Li, Dayiheng Liu, Fei Huang, Haoran Wei, Huan Lin, Jian Yang, Jianhong Tu, Jianwei Zhang, Jianxin Yang, Jiaxi Yang, Jingren Zhou, Junyang Lin, Kai Dang, Keming Lu, Keqin Bao, Kexin Yang, ...

    2025 arXiv

  11. [19]

    Colin Raffel, Noam Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, Michael Matena, Yanqi Zhou, Wei Li, and Peter J. Liu. 2020. http://jmlr.org/papers/v21/20-074.html Exploring the limits of transfer learning with a unified text-to-text transformer . Journal of Machine Lea...

    2020

  12. [20]

    Nils Reimers and Iryna Gurevych. 2019. Sentence-bert: Sentence embeddings using siamese bert-networks. In Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJC...

    2019

  13. [21]

    Liang Shi, Zhengju Tang, and Zhi Yang. 2024. A survey on employing large language models for text-to-sql tasks. arXiv preprint arXiv:2407.15186

    2024 arXiv

  14. [22]

    Michael Staniek, Raphael Schumann, Maike Z \"u fle, and Stefan Riezler. 2024. https://doi.org/10.1162/tacl_a_00654 Text-to- O verpass QL : A natural language interface for complex geodata querying of O pen S treet M ap . Transactions of the Association for Computational Lingui...

    2024 doi

  15. [23]

    Zhi Rui Tam, Cheng-Kuang Wu, Yi-Lin Tsai, Chieh-Yen Lin, Hung-yi Lee, and Yun-Nung Chen. 2024. https://doi.org/10.18653/v1/2024.emnlp-industry.91 Let me speak freely? a study on the impact of format restrictions on large language model performance. In Proceedings of the 2024 C...

    2024 doi

  16. [24]

    Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothée Lacroix, Baptiste Rozière, Naman Goyal, Eric Hambro, Faisal Azhar, Aurelien Rodriguez, Armand Joulin, Edouard Grave, and Guillaume Lample. 2023. https://arxiv.org/abs/2302.13971 Llama:...

    2023 arXiv

  17. [25]

    Overpass Turbo. 2025. https://overpass-turbo.eu/ Overpass turbo website . Accessed: 2025-02-10

    2025

  18. [26]

    Simon Will. 2021. Nlmaps web: A natural language interface to openstreetmap. In Proceedings of the Academic Track, State of the Map 2021, pages 13--15

    2021

  19. [27]

    Chao Zhang, Yuren Mao, Yijiang Fan, Yu Mi, Yunjun Gao, Lu Chen, Dongfang Lou, and Jinshu Lin. 2024. https://doi.org/10.1145/3626246.3653375 Finsql: Model-agnostic llms-based text-to-sql framework for financial analysis . In Companion of the 2024 International Conference on Man...

    2024

  20. [28]

    Xiaohu Zhu, Qian Li, Lizhen Cui, and Yongkang Liu. 2024. Large language model enhanced text-to-sql generation: A survey. arXiv preprint arXiv:2410.06011

    2024 arXiv

  21. [29]

    online" 'onlinestring :=

    ENTRY address archivePrefix author booktitle chapter edition editor eid eprint eprinttype howpublished institution journal key month note number organization pages publisher school series title type volume year doi pubmed url lastchecked label extra.label sort.label short.list...

  22. [30]

    write newline

    " write newline "" before.all 'output.state := FUNCTION n.dashify 't := "" t empty not t #1 #1 substring "-" = t #1 #2 substring "--" = not "--" * t #2 global.max substring 't := t #1 #1 substring "-" = "-" * t #2 global.max substring 't := while if t #1 #1 substring * t #2 gl...

Pith tools

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