Pith. sign in

REVIEW 4 major objections 4 minor 1 cited by

WILDCHAT-50M: A Deep Dive Into the Role of Synthetic Data in Post-Training

T0 review · 4 major / 4 minor · reviewed 2026-08-09 · deepseek-v4-flash

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

Pith's one-line read The paper introduces WILDCHAT-50M, a 50-million-conversation synthetic chat corpus generated by 54 open-weight model variants, and shows that a three-source SFT mix built from the best generator's responses outperforms the Tulu-3 mixture…

desk verdict The dataset is a genuine step-change resource; the RE-WILD superiority claim is plausible but currently rides on a spider chart and unreported generation settings. read the letter →

arxiv 2501.18511 v2 pith:J7IYASYI submitted 2025-01-30 cs.LG cs.CL

Benjamin Feuer , Chinmay Hegde This is my paper · ORCID
classification cs.LGcs.CL
keywords syntheticdatapost-trainingsupervisedfine-tuninggeneratingmodelschattranscriptsdatasetreleaseLLMevaluationcuration
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

This paper tries to establish that the choice of model generating synthetic training data is a decisive factor in how well a fine-tuned language model performs, and that a large public corpus of such data can beat carefully curated mixes. The authors build WILDCHAT-50M, the largest public chat dataset to date, by having over 50 open-weight models respond to real user prompts from the WildChat corpus. Fine-tuning Llama-3.1-8B on a three-source mix of that data, called RE-WILD, outperforms the larger Tulu-3 SFT mixture on nine post-training benchmarks while using only 40% of its samples. If correct, the paper gives the field a public resource for studying synthetic data quality and evidence that modest, well-chosen synthetic sets can displace larger curated ones.

What carries the argument

The central object is the dataset itself: WILDCHAT-50M, roughly 50 million chat conversations in which real user prompts from the existing WildChat corpus are re-answered by 54 open-weight model variants ranging from 0.5B to 104B parameters. The comparative method is synthetic data quality (SDQ) measurement: each generating model's responses become a supervised fine-tuning set for a fixed base model (Llama-3.1-8B), and the resulting fine-tunes are scored on a standardized nine-benchmark suite mixing ground-truth and LLM-judge evaluations. The RE-WILD mix then combines 246,750 conversations from the best-performing generator with 99,800 MMLU auxiliary items and 20,000 math items, showing that a small, complementary-skills blend can beat a larger curated mixture.

What would settle it

Re-run the six-generator comparison of Table 2 with identical generation settings, including the same temperature, the same context truncation, and the same token limits, across all generators; if the ranking of teachers by downstream SFT performance changes materially, the paper's assignment of credit to DGM identity is confounded by generation configuration. A second check is to reproduce RE-WILD's advantage over the Tulu-3 SFT mixture using only the officially released hyperparameters and an independent random seed.

Watch

Extended reading notes

Core claim

The paper's central discovery is that synthetic data quality is driven less by which benchmark a generating model is good at and more by domain-agnostic response traits, and that these traits are inherited during supervised fine-tuning. Across six data-generating models from four model families, ranging from 0.5B to 104B parameters, no single model dominates nine benchmarks, and parameter count does not predict data quality. Fine-tuned models track their teacher's stylistic choices, such as structure, comprehensiveness, and tone, almost exactly, while skills such as world knowledge and mathematics transfer only when the data is specifically curated for them. The paper also reports that blending responses from multiple generating models gives no benefit over single-model data, and that responses from diverse LLMs resemble each other far more than human responses to the same prompts would.

Load-bearing premise

The paper assumes that differences in downstream model performance are caused by which model generated the training data, even though generation settings were not standardized: context windows were tuned per model, and temperature values, which the paper's own appendix says substantially influence output diversity and quality, were never reported.

Editorial extensions

If this is right

  • Researchers can replace hand-curated SFT mixtures with response data from a single strong open-weight model and match or beat larger blends at lower cost.
  • DGM selection, not prompt diversity, is the main lever for generalist chat quality; blending multiple response generators adds nothing beyond averaging their performance.
  • Style traits such as comprehensiveness and formatting are inherited by fine-tunes, so dataset builders can control the stylistic fingerprint of a model by choosing which generator answers the prompts.
  • Scaling synthetic data improves SFT performance for some generators but tapers off quickly for others, so the value of collecting more data depends heavily on which model generated it.
  • Fine-tuned models do not inherit their teacher's benchmark-specific strengths, so a good teacher is not necessarily one that scores well on the target benchmark.

Reading between the lines

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

  • A testable extension the paper leaves implicit: running the same generation pipeline at fixed temperature and context settings across all 50 models would separate DGM quality from generation configuration; the paper's own appendix warns that temperature settings substantially influence output diversity and quality, yet never reports the values used.
  • The finding that response blends benchmark between their parts suggests that the prompt-diversity benefits established for large real datasets do not transfer to response space, implying that curation effort should concentrate on selecting teachers rather than mixing their outputs.
  • The styling-inheritance result suggests a cheap diagnostic for synthetic data quality: measure structural traits of a candidate generator's outputs before committing compute to full-scale generation.
  • The near-chance agreement between teacher and student benchmark strengths implies that distillation rankings can shift at inference time; users of such fine-tunes may want to re-evaluate models trained on their own outputs rather than assuming student rankings mirror teachers.
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 / 4 minor

Summary. The paper introduces WILDCHAT-50M, a large corpus of synthetic chat transcripts generated by 54 open-weight data-generating models (DGMs) on prompts from WildChat. It reports analyses of generation efficiency, intra-LLM response similarity, and the effect of DGM choice on downstream SFT performance, and introduces RE-WILD, a three-source SFT mix that is claimed to outperform the Tulu-3 SFT mixture with only 40% as many samples. The main empirical claims are that DGM identity strongly and unpredictably affects downstream SFT quality, that RE-WILD is a strong SFT mix, and that several secondary findings hold: data scaling helps, blending DGMs does not, style is inherited from the DGM, and on-policy DGMs are more effective teachers. The paper also provides the dataset, samples, and code in a public repository.

Significance. If the claims hold, the paper would provide a valuable public artifact: a chat post-training corpus far larger than prior public chat sets, together with evidence about the sensitivity of SFT to data-generating-model choice. The authors release the dataset, samples, and code, and the evaluation covers nine benchmarks with a standardized framework. The strongest assets are the scale and diversity of the generated corpus and the extensive benchmarking effort. However, the comparative conclusions are currently under-supported because generation settings are neither controlled nor reported, and the headline RE-WILD comparison is not numerically documented in the manuscript itself.

major comments (4)
  1. [§2.1, Appendix F] The DGM comparison in Table 2 and the choice of Qwen-2.5-72B as the main RE-WILD component are confounded by unstandardized and unreported generation settings. Section 2.1 reports that context windows were 'heuristically maximize[d]' and range from 2,048 to 20,000 tokens, and that the largest models were run in FP8 quantization, while sampling temperatures are never reported. Appendix F states that temperature settings 'substantially influence output diversity and quality' and recommends that researchers 'Meticulously document all hyperparameters used in generation'—a recommendation the paper itself does not follow. Since Section 3.2 demonstrates that SFT models inherit stylistic properties from their DGMs, differences in response length and style across DGMs could drive part of the Table 2 ranking. The context-truncation ablation in Section 3.2 tests only one model and only the context-length dimension, so it cannot rule out the confound. The central comparative claims require either standardized generation settings or an explicit sensitivity analysis that varies temperature and context length while holding the DGM fixed.
  2. [§3.1, Fig. 1] The headline claim that RE-WILD 'outperforms the recent Tulu-3 SFT mixture from Allen AI with only 40% as many samples' is not supported by any numeric table in the manuscript: Figure 1 is a spider chart and the text directs readers to the GitHub repository for exact scores. No confidence intervals are reported for the RE-WILD versus baseline comparison. In addition, Section 3.1 says RE-WILD components were chosen 'heuristically' while Section 3.3 describes the 'empirical process by which we determined which DGMs had high SDQ'; because the same benchmarks are used for both selection and evaluation, the authors should discuss this selection bias and should include the numeric breakdown in the paper itself.
  3. [§3.2, App. B, Table 6] The claim that fine-tuning on on-policy or same-family DGMs 'produces stronger benchmark results' is contradicted by the paper's own Table 6. For the Llama-3.1-8B target, L8B:L8I achieves Avg 0.36, identical to L8B:L70 and only 0.01 above L8B:Q7; the Table 6 caption's statement that 'Both Llama and Qwen benefit from more similar upstream models' is therefore not supported for Llama. The main-text conclusion should be restricted to the Qwen target or supported with additional runs and significance estimates.
  4. [§2.2, §3.2, Figs. 1-2] The central quantitative findings rest on single training runs without seeds or run-to-run variance. Table 2 reports 95% confidence intervals for some benchmark columns but not for MTBench or AlpacaEval, and Figure 2's scaling curves and Figure 1's spider chart have no error bars. Because differences in Avg as small as 0.01 (e.g., L8B:L8I vs L8B:L70 in Table 6) are used to draw conclusions, the paper should either add repeated-seed results or explicitly state that differences of this size are within noise.
minor comments (4)
  1. [§1, §2.2] The paper alternates between '50 different open-weight models', '54 DGMs', 'WILDCHAT-50M', and 'over 125 million chat transcripts in aggregate'; please clarify whether the dataset unit is conversations, responses, or turns and make the counts consistent.
  2. [§3.2, Fig. 2] There are several typos: 'attactive' should be 'attractive', and 'A vgis' in the Figure 2 caption should be 'Avg is'.
  3. [Appendix F] The symbols '¡' and '¿' before 0.3 and 1.2 should be '<' and '>'.
  4. [§3.3] The sentence 'On the AlpacaEval leaderboard, Qwen 2.5-72B and Llama 3.3-70B are essentially tied; but Qwen 2.5-72B is a superior DGM (as measured by AlpacaEval performance)' is self-contradictory and should be rephrased.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity found: central comparisons are external benchmark measurements, with non-load-bearing self-citations and a documented generation-setting confound.

full rationale

The paper's derivation chain is empirical rather than definitional. WILDCHAT-50M is constructed by generating responses from 54 DGMs on WildChat prompts; the DGM-quality analysis (Table 2) fine-tunes a fixed Llama-3.1-8B base on equal-size samples and scores the resulting checkpoints on standardized external benchmarks. No fitted parameter from this analysis is reused as a prediction: RE-WILD is a fixed, disclosed mix (Table 1) whose components are stated to be 'chosen heuristically to emphasize complementary skillsets,' and its comparison is against the independent Tulu-3 SFT, Magpie, and UltraChat baselines. The only self-citations (Feuer et al. 2024a,b) motivate the inclusion of both ground-truth and LLM-judge benchmarks and supply some DGM checkpoints; they do not import an unverified theorem or forbid alternatives. Two concerns do not rise to circularity: (i) RE-WILD's dominant component was identified as high-SDQ within the same paper, and the later evaluation overlaps that benchmark family, a selection-on-evaluation risk rather than a constructional identity; and (ii) Appendix F admits that unreported temperature settings 'substantially influence output diversity and quality,' which is a validity threat to the DGM-identity attribution, not a circular step because DGM identity is not defined via the downstream scores. No equation, fitted parameter, or self-citation chain makes the results equal to their inputs by construction.

Assumptions & free parameters 3 free parameters · 4 assumptions · 0 invented entities

The paper's comparative claims rest on treating DGM outputs as the only varying factor. In fact, the generation pipeline was tuned per model (context window, GPU count, quantization for large models), and the acknowledged temperature knob was not documented. The RE-WILD mix itself introduces hand-picked component sizes and was selected using benchmark feedback. No new theoretical entities are introduced.

free parameters (3)
  • RE-WILD component sizes = 246,750 / 99,800 / 20,000 conversations
    Table 1. The mix was 'chosen heuristically to emphasize complementary skillsets', i.e., tuned to benchmark performance; the sizes are not derived from a stated selection principle.
  • Per-model context window length = 2,048 to 20,000 tokens
    Section 2.1 states context windows were heuristically maximized per GPU count; this varies across DGMs and may confound comparisons of downstream SFT performance.
  • Generation sampling temperature = not reported
    Appendix F states temperature substantially affects output diversity and quality, but the data collection section does not report the temperature(s) used for any model.
assumptions (4)
  • domain assumption Open-weight Instruct model outputs generated with vLLM are representative of synthetic chat data for SFT post-training.
    The entire study treats DGM responses as training data; no human-quality baseline or closed-model comparison is used to validate this representation.
  • domain assumption The benchmark suite (MixEval, AlpacaEval2, MTBench with GPT-4o-mini, IFEval, MMLU-Pro, BBH, GPQA, MATH, MUSR) adequately measures post-training quality.
    Section 3.1 says the suite is 'standardized, balanced and large', but the authors acknowledge it omits coding and legal reasoning; the RE-WILD vs Tulu-3 claim is stated relative to this suite.
  • ad hoc to paper Selecting RE-WILD components using benchmark feedback does not invalidate the comparison to Tulu-3.
    Section 3.1: 'The datasets in this composition were chosen heuristically'. Because the same benchmarks are used for selection and evaluation, the comparison is benchmark-fitted.
  • domain assumption LLM-as-a-judge preferences with GPT-4o-mini track human preferences closely enough for the conclusions.
    Section 3.1 uses GPT-4o-mini for MTBench; Section 3.3 reasons about LLM judge criteria. Prior work on judge biases is cited, but no human validation is performed here.

how reviews work

0 comments
Cite this review

Pith. "Pith review of WILDCHAT-50M: A Deep Dive Into the Role of Synthetic Data in Post-Training." pith.science (2026). https://pith.science/paper/J7IYASYI

@misc{pith2026250118511,
  author       = {Pith},
  title        = {Pith review of: WILDCHAT-50M: A Deep Dive Into the Role of Synthetic Data in Post-Training},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/J7IYASYI}},
  note         = {Machine review of arXiv:2501.18511}
}
read the original abstract

Language model (LLM) post-training, from DPO to distillation, can refine behaviors and unlock new skills, but the open science supporting these post-training techniques is still in its infancy. One limiting factor has been the difficulty of conducting large-scale comparative analyses of synthetic data generating models and LLM judges. To close this gap, we introduce WILDCHAT-50M, the largest public chat dataset to date. We extend the existing WildChat dataset to include responses not only from GPT, but from over 50 different open-weight models, ranging in size from 0.5B to 104B parameters. We conduct an extensive comparative analysis and demonstrate the potential of this dataset by creating RE-WILD, our own public SFT mix, which outperforms the recent Tulu-3 SFT mixture from Allen AI with only 40% as many samples. Our dataset, samples and code are available at https://github.com/penfever/wildchat-50m.

Figures

Figures reproduced from arXiv: 2501.18511 by the authors.

Figure 1
Figure 1. RE-WILD outperforms strong baselines, on average, across nine benchmarks. In particular, it exhibits strong performance on generalist chat and instruction following benchmarks. MT Bench scores here are divided by 10, so that the scale is similar to our other evaluations. For the exact numeric scores for all models, please refer to our GitHub repository. Figure best viewed in color. Model MTBench AlpacaEval BBH GPQA … view at source ↗
Figure 2
Figure 2. Data scaling improves SFT performance. The effect is, however, somewhat dependent on SDQ – for DGMs such as GPT 3.5, the benefits taper off relatively quickly, but for the other three DGMs we consider, they continue to increase. Avg is the average performance over (MixEval, AlpacaEval2-LC, MTBench / 10, OpenLLM LB 2). How much does the choice of data generating model im￾pact downstream performance? Can we be certain… view at source ↗
Figure 3
Figure 3. Key words more common in L8B : L70 judgments. The more negative tone of these judgments emphasizes words like clearer (as in, “could have been clearer”), lacks, convoluted and repetitive. From these results, it seems that SDQ on a set of generalist 7 [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗
Figures from the paper (1 more)
Figure 4
Figure 4. Figure 4: Key words more common in L8B : Q72 judgments. These judgments tended to be more positive; emphasis was placed on words like appropriate, necessary, comprehensive and accurate. Model MTBench PrefRt-L8L PrefRt-L8Q PrefRt-L70B PrefRt-Q72B FlipCt L8B : L70 6.38 N/A 22.5 8.…

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

  1. Auditing LLM Editorial Bias in News Media Exposure

    cs.CY 2025-10 conditional novelty 6.0 of 10

    Compared with Google News, GPT-4o-Mini, Claude-3.7-Sonnet, and Gemini-2.0-Flash surface fewer unique news outlets, distribute attention more unevenly, and lean ideologically in system-specific ways.

Reference graph

Works this paper leans on

27 extracted references · 22 canonical work pages · cited by 1 Pith paper

  1. [1]

    Water, ammonia, and methane ices form, like the frozen tears of the damned, suspended in a sea of liquid hydrogen and helium

    Hydrogen (H₂) : The primary component of the gas giant's atmosphere, hydrogen makes up approximately succumbing to the crushing force, and heavier elements such as carbon, nitrogen, and oxygen beginning to assert their dominance. Water, ammonia, and methane ices form, like the frozen tears of the damned, suspended in a sea of liquid hydrogen and helium. T...

    work page

  2. [2]

    Accessed: 2025-01-13

    https://huggingface.co/spaces/ open-llm-leaderboard/blog, 2024. Accessed: 2025-01-13. Ganesan, K. Rouge 2.0: Updated and improved mea- sures for evaluation of summarization tasks, 2018. URL https://arxiv.org/abs/1803.01937. Gao, L., Tow, J., Abbasi, B., Biderman, S., Black, S., DiPofi, A., Foster, C., Golding, L., Hsu, J., Noac’h, A. L., Li, H., McDonell,...

    arXiv 2024

  3. [3]

    The methane molecules absorb red light, leaving the atmosphere with a predominantly blue and purple coloration

    Methane (CH₄) : Methane, present in lesser amounts, gives the planet its deep blue and purple hues. The methane molecules absorb red light, leaving the atmosphere with a predominantly blue and purple coloration. In the upper layers, methane condenses into clouds, contributing to the gothic, stormy appearance

    work page

  4. [4]

    These clouds are often tinged with dark red and black, creating a moody and dramatic effect

    Ammonia (NH₃) : Ammonia is present in the form of ice particles and clouds, adding to the planet's dark, stormy appearance. These clouds are often tinged with dark red and black, creating a moody and dramatic effect

    work page

  5. [5]

    Water ice and vapor form high- altitude clouds that can appear as ghostly, translucent layers, enhancing the gothic aesthetic

    Water (H₂O) : Water vapor is also present, contributing to the planet's atmospheric complexity. Water ice and vapor form high- altitude clouds that can appear as ghostly, translucent layers, enhancing the gothic aesthetic

    work page

  6. [6]

    how generic

    URL https://api.semanticscholar. org/CorpusID:270391432. Yasunaga, M., Shamis, L., Zhou, C., Cohen, A., Weston, J., Zettlemoyer, L., and Ghazvininejad, M. Alma: Align- ment with minimal annotation, 2024. URL https: //arxiv.org/abs/2412.04305. Zhao, W., Ren, X., Hessel, J., Cardie, C., Choi, Y ., and Deng, Y . Wildchat: 1m chatgpt interaction logs in the w...

    arXiv 2024

  7. [7]

    During these storms, the phosphine reacts with other chemicals, producing flashes of electric blue and green, adding to the eerie, gothic ambiance

    Phosphine (PH₃) : Phosphine is a trace gas that can create spectacular, bioluminescent clouds and lightning storms. During these storms, the phosphine reacts with other chemicals, producing flashes of electric blue and green, adding to the eerie, gothic ambiance

    work page

  8. [8]

    It contributes to the overall density and opacity of the atmosphere, enhancing the gothic, shadowy appearance

    Helium (He) : The second most abundant element, helium comprises about 9% of the atmosphere. It contributes to the overall density and opacity of the atmosphere, enhancing the gothic, shadowy appearance

    work page

Show all 27 references
  1. [12]

    sulfide (H₂S) are minor but significant components

    Sulfur Compounds : Sulfur dioxide (SO₂) and hydrogen Colors: - Deep indigo (#3B3F54) to sickly yellowish-green (#8B9467) - Faint, glowing blue-green (#34A8FF) of luminescent decay - Hellish, ember-like glow of the core (#FF9900) - Eerie, greenish-blue lightning flashes (#33CCF...

  2. [14]

    welfare queen

    Metallic Hydrogen : Deep within the planet, under immense pressure and temperature, hydrogen exists in a metallic state. This metallic hydrogen is responsible for the planet's strong magnetic field and contributes to its overall energy and dynamism. Additional Features • Rings...

    2023

  3. [15]

    Landowners were required to pay a proportion of their crop yields, usually in kind (grain) or in silver

    Land Tax (Diding) : The primary source of revenue for the Qing government was the land tax, which was based on the amount of land owned and its productivity. Landowners were required to pay a proportion of their crop yields, usually in kind (grain) or in silver

  4. [16]

    This tax was used to finance local government expenses, such as maintaining irrigation systems, roads, and public buildings

    Head Tax (Ding) : The head tax was a poll tax levied on all adult males, regardless of their social status or occupation. This tax was used to finance local government expenses, such as maintaining irrigation systems, roads, and public buildings

  5. [17]

    NexusFlow AI

    https://mistral.ai/news/ministraux/. NexusFlow AI. Athene: Pioneering the Future of AI Workflows, 2025. URL https://nexusflow.ai/ blogs/athene. Accessed: 2025-01-29. 11 WildChat-50M Ni, J., Xue, F., Yue, X., Deng, Y ., Shah, M., Jain, K., Neubig, G., and You, Y . Mixeval: Deri...

    2025 arXiv

  6. [18]

    Salt merchants were required to The General Secretary, like other citizens, would have paid whatever taxes were required by law

    Salt Tax : The Qing government monopolized the salt industry and imposed a tax on salt production and distribution. Salt merchants were required to The General Secretary, like other citizens, would have paid whatever taxes were required by law. However, due to their high-ranki...

    1912

  7. [19]

    Land Tax ( ⽥賦 , tián fù) : This was the primary source of purchase salt licenses from the government, which generated significant revenue

  8. [20]

    Customs Duties : The Qing government collected customs duties on imported and exported goods, including tea, silk, and other luxury items

  9. [21]

    Commercial Taxes : Merchants and traders were required to pay taxes on their businesses, including permits, licenses, and sales taxes

  10. [22]

    Administration : The taxation system was administered through a network of provincial, prefectural, and county-level governments, as well as local magistrates and tax collectors

    Tribute : The Qing government received tributes from vassal states, such as Mongolia, Tibet, and Korea, in the form of goods and services. Administration : The taxation system was administered through a network of provincial, prefectural, and county-level governments, as well ...

  11. [23]

    Ding-Yin Zhibiao

    Poll Tax ( 丁稅 , dīng shuì) : This was a per-capita tax levied on adult males. It was initially a fixed amount, regardless of wealth or social status. However, the poll tax was often criticized for being regressive and unfair to the poor. In the 18th century, the Yongzheng Empe...

  12. [24]

    The government maintained a monopoly on salt production and distribution, and merchants who sold salt had to purchase a government- issued license ( 鹽引 , yán yǐn)

    Salt Tax ( 鹽稅 , yán shuì) : The salt tax was a significant revenue source. The government maintained a monopoly on salt production and distribution, and merchants who sold salt had to purchase a government- issued license ( 鹽引 , yán yǐn). The salt tax was often a source of cor...

  13. [25]

    Various fees were levied on goods at marketplaces, ports, and along major trade routes

    Commercial Taxes ( 商稅 , shāng shuì) : These included taxes on trade and commerce. Various fees were levied on goods at marketplaces, ports, and along major trade routes. The Qing government also imposed customs duties on imported efficiency. However, the system remained largel...

    1912

  14. [26]

    Local governments often imposed additional levies to fund local projects and administration

    Miscellaneous Taxes : There were numerous other taxes and fees, including those on property transactions, funerals, and even the sale of livestock. Local governments often imposed additional levies to fund local projects and administration

  15. [27]

    This system was gradually phased out and replaced with monetary payments

    Corvée Labor ( 徭役 , yáo yì) : Although not a monetary tax, corvée labor required peasants to provide labor for public works projects, such as building roads, canals, and fortifications. This system was gradually phased out and replaced with monetary payments

  16. [28]

    The late 19th century saw increasing pressure on the tax system due to the need to fund military modernization and other reforms

    Reforms and Challenges : The Qing government faced significant challenges in collecting taxes, especially in times of natural disasters, rebellion, and economic hardship. The late 19th century saw increasing pressure on the tax system due to the need to fund military moderniza...

  17. [173]

    acl-main.173/

    URL https://aclanthology.org/2020. acl-main.173/. Meng, Y ., Xia, M., and Chen, D. Simpo: Simple preference optimization with a reference-free reward. arXiv preprint, 2024. Mishra, S., Khashabi, D., Baral, C., and Hajishirzi, H. Cross- task generalization via natural language ...

    2020

  18. [244]

    https://aclanthology.org/2022.acl- long.244

    URL, May 2022. https://aclanthology.org/2022.acl- long.244. Mistral, A. I. Ministraux: Pushing the boundaries of effi- cient transformer design. URL, 2024. Retrieved 2024-11-

    2022

  19. [2024]

    https://zenodo.org/records/12608602

    URL, 2024. https://zenodo.org/records/12608602. GLM, T., :, Zeng, A., Xu, B., Wang, B., Zhang, C., Yin, D., Zhang, D., Rojas, D., Feng, G., Zhao, H., Lai, H., Yu, H., Wang, H., Sun, J., Zhang, J., Cheng, J., Gui, J., Tang, J., Zhang, J., Sun, J., Li, J., Zhao, L., Wu, L., Zhon...

    2024

Pith tools

Reviewed August 9, 2026 · model on record in the stance chip above.