Pith. sign in

REVIEW 2 major objections 5 minor 161 references

DFM Mimir v1: An Open HRM Delivering Frontier Performance at 1B Parameters Using Only Permissible Post-Training Data

T0 review · 2 major / 5 minor · reviewed 2026-08-14 · deepseek-v4-flash

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

Pith's one-line read Mimir v1, a 1-billion-parameter model trained from scratch on 161 permissible post-training datasets, outperforms its 1B predecessor and matches or beats models two to five times larger, setting a new Danish state of the art.

desk verdict Mimir v1 is a real open-data model with careful documentation, but the Danish SOTA claim is unsecured by apparent train/eval overlap and needs a decontamination analysis before it can be trusted. read the letter →

arxiv 2608.13517 v1 pith:64KO2M5S submitted 2026-08-13 cs.CL cs.AI

Peter Schneider-Kamp , Jacob Nielsen , Gianluca Barmina , Kenneth Enevoldsen , Lukas Galke Poech This is my paper · ORCID
classification cs.CLcs.AI
keywords largelanguagemodelsHRMarchitecturepermissibledataDanishbenchmarksinstructiontuningsyntheticlow-resourcelanguagesfrom-scratchtraining
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

Mimir v1 is a 1-billion-parameter language model trained from scratch using only permissible post-training data: openly licensed sources, agreement-supplied data, and EU text-and-data-mining-excepted material. The paper claims this model outperforms the original HRM-Text 1B, competes with larger models such as Qwen 3.5 4B and Gemma 4 E2B across 20 English, math, code, and Danish benchmarks, and sets a new state of the art for Danish. If the claim holds, small research groups and low-resource language communities can build competitive models without massive, often non-permissible pretraining corpora.

What carries the argument

The mechanism is the HRM-Text architecture: a transformer with 32 layers, hidden size 1536, 12 attention heads per layer, and hierarchical reasoning configured as 2 H-cycles and 3 L-cycles, trained with truncated backpropagation over five steps. This structure is paired with a data recipe built from 161 permissible datasets, including synthetic 'transplant' datasets generated with a larger teacher model and audited before inclusion. The H/L-cycle design is what lets the model learn reasoning behavior from instruction and post-training data without a conventional massive pretraining corpus, and the synthetic transplants are what let the authors replace non-permissible components of the original HRM-Text corpus while keeping the task mix intact.

What would settle it

Compare the evaluation splits of giannor/dala, giannor/gec dala tv2r it, and oliverkinch/multi-wiki-qa against the training versions of those datasets; if any evaluation query or target appears in the training data, the reported Danish scores are inflated by memorization and the state-of-the-art claim fails for those benchmarks.

Watch

Extended reading notes

Core claim

The central claim is that the Hierarchical Reasoning Model (HRM) architecture makes it possible to train a competitive 1B model from scratch on post-training data alone, provided the data is diverse and permissible. Mimir v1 is trained on 161 datasets totaling 70.48 billion tokens per epoch, with a corpus that is 68.6% English and 24.7% Danish, and it uses synthetic 'transplant' datasets to replace the non-permissible portions of the original HRM-Text data. In the paper's evaluations, Mimir v1 exceeds HRM-Text 1B on nearly every benchmark, leads the 1B weight class on BoolQ, Winogrande, DROP, GSM8K, and HumanEval, trails Qwen 3.5 4B by only 0.3 points on the English average, and records the best Danish scores on DaLA, GEC, WikiQA, and the overall Danish suite.

Load-bearing premise

The Danish state-of-the-art claim presumes that the evaluation benchmarks measure generalization rather than memorization, which requires that their test items were excluded from the training corpus; the paper does not state that the overlapping DaLA, GEC-DaLA, and Multi Wiki QA datasets were held out.

Editorial extensions

If this is right

  • Because Mimir v1 is trained only on permissible data and released openly, it provides a licensing-clean 1B base model for Danish and English applications.
  • The reported improvement over HRM-Text 1B is large, including a 36.7% gain on the Math & Code average, showing that the HRM recipe itself, not the original data, drives much of the performance.
  • The synthetic transplant strategy gives data teams a legal path to recreate proprietary-style training mixes from sources that are openly licensed or audited.
  • With about 1.65 million training steps on eight accelerators in under three weeks, the recipe is feasible for national and university-scale projects rather than only large industrial labs.

Reading between the lines

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

  • Re-running the Danish suite on evaluation splits provably disjoint from the training corpus would settle whether the reported state-of-the-art scores reflect generalization rather than memorization, since three Danish evaluation sets share identifiers with training datasets.
  • Because 68.5% of the training tokens are English, the Danish-first performance may rely on transfer from English; a Danish-only or Danish-heavy training ablation would reveal how much of the Danish score comes from the 24.7% Danish share.
  • The model adopts another family's tokenizer and chat template, so the architecture's unique contribution is not isolated; an ablation with a different tokenizer would clarify what HRM's H/L-cycle design adds on its own.
  • Extending the same data recipe to 2-4B parameters is a natural next test; the 1B results suggest such a model would surpass Qwen 3.5 4B on the English average and widen the Danish lead.
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

2 major / 5 minor

Summary. The paper presents Mimir v1, a 1-billion-parameter model based on the HRM-Text architecture, trained from scratch on a mixture of 161 datasets totaling about 70.5B tokens per epoch, all claimed to be permissible post-training data. The authors report benchmark results for English, Math & Code, and Danish, claiming that Mimir outperforms the original HRM-Text 1B, competes with larger models such as Qwen 3.5 4B and Gemma 4 E2B, and sets a new state of the art for Danish. The appendices list the complete training corpus and evaluation datasets, and the model is released on the Hugging Face Hub.

Significance. If the reported results are valid, the contribution is significant: it would demonstrate that a 1B model trained entirely on permissible data can be competitive with models several times its size, and it would provide a useful open recipe for low-resource language modeling. The paper is transparent about the training corpus, hyperparameters, and model release, and the synthetic transplant idea is interesting. However, the central evidence consists entirely of benchmark tables, and those tables are undermined by the apparent overlap between training datasets and evaluation datasets. Because the Danish state-of-the-art claim rests on benchmarks whose exact or near-identical sources appear in the training list, the results cannot currently be interpreted as evidence of generalization. The absence of error bars or decontamination analysis further weakens the headline claims.

major comments (2)
  1. [§2 Table 10; §5 Table 11] The training corpus in Table 10 contains entries whose Hugging Face identifiers match or are near-identical to evaluation datasets in Table 11. Specifically, row #43 (giannor/gec dala tv2r it) and row #61 (giannor/dala tv2r it) correspond to the GEC-DaLA (giannor/dala gen v3) and DaLA (giannor/dala) evaluation sets; row #66 (oliverkinch/multi-wiki-qa-high-quality-subset) is a subset of the Multi Wiki QA evaluation set (oliverkinch/multi-wiki-qa); and row #91 (ccdv/govreport-summarization) is a summarization variant of the GovReport evaluation source (ccdv/govreport). The manuscript never states that these evaluation splits were excluded from training, and it provides no decontamination analysis. Because the Danish state-of-the-art claim depends on DaLA, GEC-DaLA, and Multi Wiki QA scores, the reported numbers may reflect memorization rather than generalization. The central claim of the paper is therefore unsupported as written; the authors must either demonstrate that these training entries are disjoint from the evaluation benchmarks or retrain without them.
  2. [§5 Evaluation Setup] No error bars, confidence intervals, or significance tests are reported for any benchmark. Several headline comparisons are close enough that sampling noise could change the ranking; for example, the English average is 69.0 for Mimir versus 69.3 for Qwen 3.5 4B, and on MATH Mimir scores 45.8 while HRM-Text scores 56.0. All numbers come from single greedy decoding runs. Without uncertainty quantification or repeated-seed evaluation, the claims to 'outperform' competitors on individual tasks and to be 'close to the best' on others are not statistically established.
minor comments (5)
  1. [Author list and References] The author list gives 'Lukas Galke Poech' while the reference to Dynaword gives 'Lukas Galke'; the spelling should be reconciled.
  2. [§4 Training] The phrase 'fitting 4 contexts of length 4096 each' is grammatically incomplete; the intended meaning is that the per-accelerator batch comprises 4 sequences of length 4096.
  3. [§2.1 and Table 1] Section 2.1 states that the Sapient repository 'bundles 107 sub-collections', while Table 1 reports 71 datasets in the Sapient mixed category; the relationship between these counts should be clarified.
  4. [§5 Evaluation Setup] The evaluation protocol differs between Mimir, evaluated with Hugging Face Transformers, and the baselines, evaluated with vLLM via Inspect AI; the authors state the results are comparable, but no side-by-side verification of the exact tables is provided.
  5. [§2 Datasets] The claim that all training data are 'permissible' is asserted at the aggregate level, but the paper does not provide per-dataset license information or a documented permissibility audit for the 161 datasets.

Circularity Check

3 steps flagged · score 7.0 of 10

Danish SOTA claim rests on evaluation sets that appear in the training corpus; no decontamination is documented, so the benchmark scores are not independent evidence.

  1. fitted input called prediction [Table 10 rows #43 and #61 vs Table 11 DaLA/GEC-DaLA; Section 5 Results]
    "giannor/gec dala tv2r it · Reformatted · 193M ... giannor/dala tv2r it · Reformatted · 68.5M ... [Table 10] ... DaLA giannor/dala 0; GEC-DaLA giannor/dala gen v3 0 [Table 11]."

    The Danish SOTA headline is computed from Table 9, where DaLA (96.1) and GEC-DaLA (85.6) are the strongest contributors. Training rows #43 and #61 are the same benchmark family under slightly different names, and the paper provides no decontamination or split-exclusion statement. With 193M + 68.5M tokens in training and only 1.65M steps, the model can easily memorize the small evaluation sets. Thus these 'predicted' scores are not independent evidence of generalization; they may simply reflect training input, making the central Danish claim circular.

  2. fitted input called prediction [Table 10 row #66 vs Table 11 Multi Wiki QA]
    "oliverkinch/multi-wiki-qa-high-quality-subset · Reformatted · 41.9M [Table 10] ... Multi Wiki QA oliverkinch/multi-wiki-qa 0 [Table 11]."

    The Multi Wiki QA evaluation set is oliverkinch/multi-wiki-qa, and the training corpus includes oliverkinch/multi-wiki-qa-high-quality-subset, a curated subset of the same source. The paper reports a WikiQA EM of 66.8 and uses it as evidence of Danish QA ability, but no exclusion of the evaluation split is documented. If the high-quality subset contains the same questions, the score is memorization rather than prediction, directly undermining the Danish SOTA claim.

1 more flagged steps
  1. fitted input called prediction [Table 10 row #91 vs Table 11 GovReport]
    "ccdv/govreport-summarization · Reformatted · 4.4M [Table 10] ... GovReport ccdv/govreport 0 [Table 11]."

    The English GovReport evaluation set ccdv/govreport is paired with a training dataset ccdv/govreport-summarization from the same repository. GovReport is included in the English average, and the paper highlights 'Mimir outperforms all considered competitors on ... DROP' but does not establish that GovReport summaries were held out. Without a decontamination check, the GovReport R1 score of 32.0 is not a trustworthy external prediction; it may be an artifact of training on the same source.

full rationale

The paper's derivation chain is not circular in its architecture choice: the HRM-Text framework and training recipe come from Wang et al. [2026], an external source, and the model is trained from scratch with its own data mixture. The circularity is concentrated in the evaluation loop. The load-bearing claim—'sets a new state of the art for Danish' and 'highly competitive performance for English'—rests on benchmark tables whose evaluation datasets appear, under near-identical Hugging Face identifiers, in the training corpus. Specifically, Table 10 includes giannor/gec dala tv2r it and giannor/dala tv2r it for the DaLA/GEC-DaLA benchmarks, oliverkinch/multi-wiki-qa-high-quality-subset for Multi Wiki QA, and ccdv/govreport-summarization for GovReport. The paper never states that these evaluation splits were excluded or deduplicated, and it contains no decontamination analysis. Since the reported scores on these benchmarks are the main evidence for the headline results, the predictions are not independent of the training input; they are, at least in part, a reflection of the training corpus itself. This is a partial but central circularity: if the overlap is confirmed, the Danish SOTA and several English results reduce to memorization rather than generalization. The remaining benchmark results (e.g., BoolQ, Winogrande, GSM8K) still provide some independent signal, so the paper is not fully circular, but the core claim is compromised without a holdout guarantee.

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

The central claim is empirical and does not rest on a mathematical derivation. The main unstated premise is that the evaluation benchmarks are held out from training. Additional assumptions include the suitability of the borrowed tokenizer and chat template, the effectiveness of the cited HRM-Text architecture, and the comparability of different evaluation harnesses. The free parameters are dataset repetition weights and acceptance thresholds for synthetic data, both chosen by hand. No new entities are introduced.

free parameters (2)
  • Dataset repetition multipliers = e.g., lærebogen 4x, 8 small Danish datasets 10x, Dolci-No-Tools 2x
    Chosen by hand to balance coverage of Danish and English instruction; not derived from principles.
  • Synthetic data acceptance thresholds = single-digit to high-nineties percent depending on category
    Selected by the authors during quality auditing; the thresholds materially shape the synthetic portion of the corpus.
assumptions (4)
  • domain assumption The Gemma-4 tokenizer and chat template are suitable for Danish text.
    The model inherits these from Gemma 4 without a dedicated Danish tokenizer validation.
  • domain assumption The HRM-Text architecture is an effective architecture for instruction-tuned training from scratch.
    The paper relies on Wang et al. (2026) for this property rather than demonstrating it independently.
  • domain assumption Evaluation benchmarks measure held-out generalization and are not included in the training data.
    This is the central premise of the benchmark claims; the paper does not test for overlap, and dataset IDs suggest overlap for DaLA, GEC, and Multi Wiki QA.
  • domain assumption Evaluating Mimir with Hugging Face Transformers and baselines with Inspect AI yields comparable scores.
    The paper notes both were run but reports HF numbers for Mimir and Inspect AI numbers for baselines without a controlled comparison.

how reviews work

0 comments
Cite this review

Pith. "Pith review of DFM Mimir v1: An Open HRM Delivering Frontier Performance at 1B Parameters Using Only Permissible Post-Training Data." pith.science (2026). https://pith.science/paper/64KO2M5S

@misc{pith2026260813517,
  author       = {Pith},
  title        = {Pith review of: DFM Mimir v1: An Open HRM Delivering Frontier Performance at 1B Parameters Using Only Permissible Post-Training Data},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/64KO2M5S}},
  note         = {Machine review of arXiv:2608.13517}
}
read the original abstract

Current large language model development relies on massive, often non-permissible datasets, creating a high barrier for researchers committed to open-source and ethically sourced data. We introduce Mimir v1, a 1-billion-parameter language model based on the Hierarchical Reasoning Model (HRM) architecture, that is trained from scratch and delivers highly competitive performance for English and sets a new state of the art for Danish using only permissible post-training data. Trained on a mixture of 161 datasets, Mimir v1 outperforms the original HRM-Text 1B and competes with larger frontier models like Qwen 3.5 4B and Gemma 4 E2B, tested across 20 benchmarks for English, Math & Code and Danish. The model is available on the Hugging Face Hub: https://huggingface.co/danish-foundation-models/DFM-Mimir

Figures

Figures reproduced from arXiv: 2608.13517 by the authors.

Figure 1
Figure 1. Aggregate results comparing DFM Mimir 1B against the HRM-Text 1B, Qwen 3.5 2B and [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

161 extracted references · 80 canonical work pages

  1. [1]

    sapientinc/HRM-Text-data-io-cleaned- 20260515 Curated + refor- matted 11.92B 16.91%

  2. [2]

    danish-foundation-models/laerebogen Reformatted 8.32B 11.81%

  3. [3]

    nvidia/OpenMathInstruct-2 Reformatted 6.60B 9.37%

  4. [4]

    nvidia/Nemotron-SFT -Agentic-v2 Reformatted 4.27B 6.06%

  5. [5]

    danish-foundation-models/dfm-dyna- instruct Reformatted 3.54B 5.03%

  6. [6]

    allenai/Dolci-Instruct-SFT -No-Tools Reformatted 3.49B 4.95%

  7. [7]

    schneiderkamplab/opus-da-en- permissive Reformatted 2.90B 4.12%

  8. [8]

    allenai/Dolci-Instruct-SFT Reformatted 2.24B 3.17%

Show all 161 references
  1. [9]

    nvidia/AceReason-1.1-SFT Reformatted 1.95B 2.76%

  2. [10]

    allenai/big-reasoning-traces Reformatted 1.66B 2.35%

  3. [11]

    allenai/Dolci-Instruct-SFT -Tool-Use Tool-call format- ted 1.61B 2.29%

  4. [12]

    nvidia/Nemotron-SFT -Instruction- Following-Chat-v2 Reformatted 1.60B 2.27%

  5. [13]

    allenai/tulu-3-sft-mixture Reformatted 1.57B 2.23%

  6. [14]

    laion/Scientific-Summaries Reformatted 1.27B 1.80%

  7. [15]

    schneiderkamplab/common-pile-prefix- continuation Synthetic + au- dited 1.15B 1.64%

  8. [16]

    synquid/wiki-instruct-da Reformatted 988M 1.40%

  9. [17]

    schneiderkamplab/dfm8-openhermes-da Translated + au- dited 922M 1.31%

  10. [18]

    schneiderkamplab/common-pile- denoising Synthetic + au- dited 884M 1.25%

  11. [19]

    allenai/tulu-v2-sft-mixture Reformatted 840M 1.19%

  12. [20]

    schneiderkamplab/common-pile-span- filling Synthetic + au- dited 822M 1.17%

  13. [21]

    schneiderkamplab/dfm8-openhermes-en Translated + au- dited 672M 0.95366%

  14. [22]

    allenai/tulu-v2-sft-long-mixture Reformatted 605M 0.85902%

  15. [23]

    allenai/verifiable-reasoning-filtered-gpt- 41 Reformatted 605M 0.85798%

  16. [24]

    open-thoughts/OpenThoughts2-1M Reformatted 534M 0.75834% 12 continued from previous page # Dataset Form Tokens/epoch Share

  17. [25]

    schneiderkamplab/dfm8-synthetic- native-tool-calling Synthetic + au- dited 513M 0.728%

  18. [26]

    schneiderkamplab/transformations- danish-danish Synthetic + au- dited 496M 0.70403%

  19. [27]

    schneiderkamplab/transformations- english-english Synthetic + au- dited 471M 0.66842%

  20. [28]

    schneiderkamplab/transformations- danish-english Synthetic + au- dited 469M 0.66529%

  21. [29]

    schneiderkamplab/transformations- english-danish Synthetic + au- dited 418M 0.59274%

  22. [30]

    nvidia/Nemotron-SFT -Multilingual-v1 Reformatted 413M 0.58565%

  23. [31]

    schneiderkamplab/dfm8-synthetic- danish-summarization-rewrite-controls Synthetic + au- dited 396M 0.56181%

  24. [32]

    MegaScience/TextbookReasoning Reformatted 374M 0.5311%

  25. [33]

    schneiderkamplab/dfm8-synthetic- multiturn-danish-english-chat Synthetic + au- dited 366M 0.51997%

  26. [34]

    DBC (agreement-supplied) Agreement- supplied 356M 0.50529%

  27. [35]

    schneiderkamplab/danish-dynaword- denoising Synthetic + au- dited 323M 0.45861%

  28. [36]

    Lex.dk articles Agreement- supplied 313M 0.44467%

  29. [37]

    oliverkinch/machine-translation-da-en Reformatted 284M 0.40348%

  30. [38]

    schneiderkamplab/danish-dynaword- prefix-continuation Synthetic + au- dited 252M 0.35728%

  31. [39]

    schneiderkamplab/danish-dynaword- span-filling Synthetic + au- dited 251M 0.35593%

  32. [40]

    schneiderkamplab/dfm8-synthetic-code- debugging Synthetic + au- dited 247M 0.34996%

  33. [41]

    GEM/wiki cat sum Reformatted 205M 0.29117%

  34. [42]

    allenai/tulu-3-sft-personas-math Reformatted 204M 0.28922%

  35. [43]

    giannor/gec dala tv2r it Reformatted 193M 0.27377%

  36. [44]

    synquid/wildchat-100k-qwen-messages Reformatted 190M 0.2694%

  37. [45]

    schneiderkamplab/dfm8-synthetic-strict- math-answer-contract Synthetic + au- dited 173M 0.24605%

  38. [46]

    schneiderkamplab/dfm8-synthetic- constrained-format-following Synthetic + au- dited 171M 0.24275%

  39. [47]

    oliverkinch/danish-summarization Reformatted 168M 0.23888%

  40. [48]

    schneiderkamplab/common-pile- paragraph-reordering Synthetic + au- dited 167M 0.23653%

  41. [49]

    schneiderkamplab/danish-dynaword- paragraph-reordering Synthetic + au- dited 160M 0.22769% 13 continued from previous page # Dataset Form Tokens/epoch Share

  42. [50]

    glaiveai/glaive-function-calling-v2 Tool-call format- ted 156M 0.22066%

  43. [51]

    oliverkinch/machine-translation-da-ar Reformatted 140M 0.19802%

  44. [52]

    allenai/SciRIFF-train-mix Reformatted 132M 0.18771%

  45. [53]

    common-pile/arxiv papers filtered Derived task 130M 0.18386%

  46. [54]

    oliverkinch/da-instruct-dynaword Reformatted 103M 0.14592%

  47. [55]

    synquid/translation-100k Reformatted 98.2M 0.13938%

  48. [56]

    kobprof/skolegpt-instruct Reformatted 86.4M 0.12259%

  49. [57]

    allenai/verifiable-reasoning-filtered-o4- mini Reformatted 80.2M 0.11378%

  50. [58]

    oliverkinch/tidsskrift-dk-bt Reformatted 76.7M 0.10878%

  51. [59]

    allenai/open math 2 50k r1-original Reformatted 71.8M 0.1019%

  52. [60]

    oliverkinch/machine-translation-da-uk Reformatted 69.5M 0.09859%

  53. [61]

    giannor/dala tv2r it Reformatted 68.5M 0.09723%

  54. [62]

    Salesforce/xlam-function-calling-60k Tool-call format- ted 67.0M 0.09502%

  55. [63]

    oliverkinch/danish-qa Reformatted 57.5M 0.08153%

  56. [64]

    oliverkinch/dst-table-prompts-bt Reformatted 51.1M 0.07248%

  57. [65]

    danish-foundation- models/ai arena udtraek Reformatted 45.7M 0.06483%

  58. [66]

    oliverkinch/multi-wiki-qa-high-quality- subset Reformatted 41.9M 0.05947%

  59. [67]

    oliverkinch/dynaword-bt Reformatted 34.3M 0.04867%

  60. [68]

    HuggingFaceH4/no robots Reformatted 31.4M 0.04459%

  61. [69]

    oliverkinch/da-instruct-dynaword- contemporary-hq Reformatted 25.6M 0.03626%

  62. [70]

    Team-ACE/ToolACE Tool-call format- ted 25.2M 0.03579%

  63. [71]

    oliverkinch/da-instruct-dynaword-hq Reformatted 24.1M 0.03414%

  64. [72]

    oliverkinch/danish-university-portals-bt Reformatted 21.8M 0.03088%

  65. [73]

    allenai/tulu-3-sft-personas-algebra Reformatted 21.0M 0.02982%

  66. [74]

    allenai/IF sft data verified Reformatted 19.9M 0.02829%

  67. [75]

    oliverkinch/autodata-da-sft Reformatted 18.9M 0.02677%

  68. [76]

    oliverkinch/danmarks-statistik-bt Reformatted 18.9M 0.02676%

  69. [77]

    synquid/danish-verifiable-reasoning Reformatted 18.1M 0.02573%

  70. [78]

    oliverkinch/da-instruct-dynaword- contemporary Reformatted 17.6M 0.02491%

  71. [79]

    oliverkinch/eur-lex-bt Reformatted 16.7M 0.02369%

  72. [80]

    allenai/tulu-3-sft-personas-code Reformatted 14.6M 0.02077%

  73. [81]

    oliverkinch/instruct-bt Reformatted 13.5M 0.0191%

  74. [82]

    schneiderkamplab/sapient-synth-flan- dialog-fsopt-data-qrecc Synthetic + au- dited 13.1M 0.0186% 14 continued from previous page # Dataset Form Tokens/epoch Share

  75. [83]

    synquid/ifbench-train Reformatted 12.7M 0.018%

  76. [84]

    allenai/tulu-3-sft-personas-instruction- following Reformatted 11.5M 0.01634%

  77. [85]

    schneiderkamplab/sapient-synth-flan- flan-fsopt-data-aeslc-1.0.0 Synthetic + au- dited 10.2M 0.01441%

  78. [86]

    synquid/mt-da-deepseek Reformatted 8.9M 0.01261%

  79. [87]

    allenai/RLVR-MATH Reformatted 8.1M 0.01145%

  80. [88]

    allenai/RLVR-GSM Reformatted 6.8M 0.00968%

  81. [89]

    allenai/Dolci-Instruct-SFT -Tool-Use-SA Tool-call format- ted 5.8M 0.00825%

  82. [90]

    schneiderkamplab/sapient-synth-flan- flan-fsnoopt-data-aeslc-1.0.0 Synthetic + au- dited 4.8M 0.0068%

  83. [91]

    ccdv/govreport-summarization Reformatted 4.4M 0.00626%

  84. [92]

    schneiderkamplab/sapient-synth-flan- flan-fsopt-data-opinion-abstracts-rotten- tomatoes Synthetic + au- dited 4.3M 0.00612%

  85. [93]

    schneiderkamplab/sapient-synth-flan- dialog-fsopt-data-qrecc-ii Synthetic + au- dited 2.9M 0.00413%

  86. [94]

    schneiderkamplab/sapient-synth-flan- niv2-fsopt-data-task589-amazonfood- summary-text-generation Synthetic + au- dited 2.8M 0.00399%

  87. [95]

    oliverkinch/eur-lex-sum-instruct Reformatted 2.6M 0.00373%

  88. [96]

    schneiderkamplab/sapient-synth-flan- niv2-fsopt-data-task590-amazonfood- summary-correction-classification Synthetic + au- dited 2.5M 0.0036%

  89. [97]

    schneiderkamplab/sapient-synth-flan- niv2-fsopt-data-task618-amazonreview- summary-text-generation Synthetic + au- dited 2.2M 0.00309%

  90. [98]

    schneiderkamplab/sapient-synth- flan-niv2-fsopt-data-task1309- amazonreview-summary-classification Synthetic + au- dited 1.9M 0.00276%

  91. [99]

    schneiderkamplab/sapient-synth-flan- flan-zsnoopt-data-aeslc-1.0.0 Synthetic + au- dited 1.9M 0.00271%

  92. [100]

    schneiderkamplab/sapient-synth-flan- flan-fsnoopt-data-opinion-abstracts- rotten-tomatoes Synthetic + au- dited 1.7M 0.00247%

  93. [101]

    schneiderkamplab/sapient-synth-flan- flan-zsopt-data-aeslc-1.0.0 Synthetic + au- dited 1.7M 0.00245%

  94. [102]

    schneiderkamplab/sapient-synth-flan- dialog-zsopt-data-qrecc Synthetic + au- dited 1.7M 0.00245% 15 continued from previous page # Dataset Form Tokens/epoch Share

  95. [103]

    schneiderkamplab/sapient-synth-flan- niv2-fsopt-data-task1375-newscomm- translation Synthetic + au- dited 1.2M 0.00166%

  96. [104]

    schneiderkamplab/sapient-synth-flan- niv2-fsopt-data-task1373-newscomm- translation Synthetic + au- dited 1.1M 0.0015%

  97. [105]

    schneiderkamplab/sapient-synth-flan- niv2-fsopt-data-task1376-newscomm- translation Synthetic + au- dited 974K 0.00138%

  98. [106]

    schneiderkamplab/sapient-synth- platypus-reclor Synthetic + au- dited 922K 0.00131%

  99. [107]

    schneiderkamplab/sapient-synth- tasksource-reclor Synthetic + au- dited 901K 0.00128%

  100. [108]

    schneiderkamplab/sapient-synth-flan- niv2-fsopt-data-task902-deceptive- opinion-spam-classification Synthetic + au- dited 878K 0.00125%

  101. [109]

    schneiderkamplab/sapient-synth-flan- niv2-fsopt-data-task634-allegro-reviews- classification Synthetic + au- dited 816K 0.00116%

  102. [110]

    schneiderkamplab/sapient-synth-flan- niv2-fsopt-data-task635-allegro-reviews- answer-generation Synthetic + au- dited 816K 0.00116%

  103. [111]

    schneiderkamplab/sapient-synth-flan- niv2-zsopt-data-task590-amazonfood- summary-correction-classification Synthetic + au- dited 811K 0.00115%

  104. [112]

    schneiderkamplab/sapient-synth- flan-niv2-zsopt-data-task1309- amazonreview-summary-classification Synthetic + au- dited 795K 0.00113%

  105. [113]

    schneiderkamplab/sapient-synth-flan- flan-fsopt-data-opinion-abstracts-idebate Synthetic + au- dited 746K 0.00106%

  106. [114]

    schneiderkamplab/sapient-synth-flan- niv2-fsopt-data-task1370-newscomm- classification Synthetic + au- dited 721K 0.00102%

  107. [115]

    schneiderkamplab/sapient-synth-flan- niv2-zsopt-data-task589-amazonfood- summary-text-generation Synthetic + au- dited 705K 0.001%

  108. [116]

    schneiderkamplab/sapient-synth-flan- niv2-fsopt-data-task1371-newscomm- translation Synthetic + au- dited 703K 0.001%

  109. [117]

    oliverkinch/doab-da-bt Reformatted 694K 0.00098% 16 continued from previous page # Dataset Form Tokens/epoch Share

  110. [118]

    schneiderkamplab/sapient-synth-flan- niv2-fsopt-data-task1377-newscomm- translation Synthetic + au- dited 679K 0.00096%

  111. [119]

    danish-foundation-models/kaenguruen Reformatted 638K 0.0009%

  112. [120]

    schneiderkamplab/sapient-synth-flan- flan-zsopt-data-opinion-abstracts-rotten- tomatoes Synthetic + au- dited 622K 0.00088%

  113. [121]

    schneiderkamplab/sapient-synth-flan- flan-zsnoopt-data-opinion-abstracts- rotten-tomatoes Synthetic + au- dited 616K 0.00087%

  114. [122]

    schneiderkamplab/sapient-synth-flan- niv2-fsopt-data-task870-msmarco- answer-generation Synthetic + au- dited 599K 0.00085%

  115. [123]

    schneiderkamplab/sapient-synth-flan- niv2-fsopt-data-task1374-newscomm- translation Synthetic + au- dited 599K 0.00085%

  116. [124]

    schneiderkamplab/sapient-synth- tasksource-pragmeval-sarcasm Synthetic + au- dited 542K 0.00077%

  117. [125]

    schneiderkamplab/sapient-synth-flan- niv2-fsopt-data-task903-deceptive- opinion-spam-classification Synthetic + au- dited 535K 0.00076%

  118. [126]

    schneiderkamplab/sapient-synth-flan- niv2-zsopt-data-task618-amazonreview- summary-text-generation Synthetic + au- dited 532K 0.00076%

  119. [127]

    schneiderkamplab/sapient-synth-flan- flan-fsnoopt-data-opinion-abstracts- idebate Synthetic + au- dited 527K 0.00075%

  120. [128]

    schneiderkamplab/sapient-synth-flan- dialog-zsopt-data-qrecc-ii Synthetic + au- dited 513K 0.00073%

  121. [129]

    schneiderkamplab/sapient-synth-flan- niv2-fsopt-data-task265-paper-reviews- language-identification Synthetic + au- dited 395K 0.00056%

  122. [130]

    schneiderkamplab/sapient-synth-flan- niv2-zsopt-data-task635-allegro-reviews- answer-generation Synthetic + au- dited 320K 0.00045%

  123. [131]

    schneiderkamplab/sapient-synth- flan-niv2-fsopt-data-task909-dialogre- prevalent-speakers Synthetic + au- dited 296K 0.00042%

  124. [132]

    schneiderkamplab/sapient-synth-flan- niv2-zsopt-data-task1376-newscomm- translation Synthetic + au- dited 274K 0.00039% 17 continued from previous page # Dataset Form Tokens/epoch Share

  125. [133]

    schneiderkamplab/sapient-synth-flan- niv2-zsopt-data-task1373-newscomm- translation Synthetic + au- dited 268K 0.00038%

  126. [134]

    schneiderkamplab/sapient-synth-flan- niv2-zsopt-data-task1375-newscomm- translation Synthetic + au- dited 265K 0.00038%

  127. [135]

    schneiderkamplab/sapient-synth-flan- niv2-fsopt-data-task266-paper-reviews- reviewer-perspective Synthetic + au- dited 236K 0.00033%

  128. [136]

    schneiderkamplab/sapient-synth- flan-niv2-fsopt-data-task906-dialogre- identify-names Synthetic + au- dited 235K 0.00033%

  129. [137]

    schneiderkamplab/sapient-synth-flan- flan-zsnoopt-data-opinion-abstracts- idebate Synthetic + au- dited 223K 0.00032%

  130. [138]

    schneiderkamplab/sapient-synth-flan- flan-zsopt-data-opinion-abstracts- idebate Synthetic + au- dited 223K 0.00032%

  131. [139]

    schneiderkamplab/sapient-synth-flan- niv2-zsopt-data-task634-allegro-reviews- classification Synthetic + au- dited 215K 0.0003%

  132. [140]

    schneiderkamplab/sapient-synth-flan- niv2-zsopt-data-task870-msmarco- answer-generation Synthetic + au- dited 212K 0.0003%

  133. [141]

    schneiderkamplab/sapient-synth-flan- niv2-zsopt-data-task902-deceptive- opinion-spam-classification Synthetic + au- dited 198K 0.00028%

  134. [142]

    schneiderkamplab/sapient-synth-flan- niv2-fsopt-data-task672-amazon-yelp- summarization Synthetic + au- dited 198K 0.00028%

  135. [143]

    schneiderkamplab/sapient-synth-flan- niv2-zsopt-data-task1377-newscomm- translation Synthetic + au- dited 182K 0.00026%

  136. [144]

    schneiderkamplab/sapient-synth-flan- niv2-zsopt-data-task1374-newscomm- translation Synthetic + au- dited 174K 0.00025%

  137. [145]

    schneiderkamplab/sapient-synth-flan- niv2-zsopt-data-task1371-newscomm- translation Synthetic + au- dited 171K 0.00024%

  138. [146]

    schneiderkamplab/sapient-synth-flan- niv2-zsopt-data-task265-paper-reviews- language-identification Synthetic + au- dited 154K 0.00022% 18 continued from previous page # Dataset Form Tokens/epoch Share

  139. [147]

    schneiderkamplab/sapient-synth-flan- niv2-zsopt-data-task903-deceptive- opinion-spam-classification Synthetic + au- dited 149K 0.00021%

  140. [148]

    schneiderkamplab/sapient-synth-flan- niv2-fsopt-data-task264-paper-reviews- accept-reject Synthetic + au- dited 140K 0.0002%

  141. [149]

    schneiderkamplab/sapient-synth- platypus-scibench Synthetic + au- dited 104K 0.00015%

  142. [150]

    schneiderkamplab/sapient-synth-flan- niv2-zsopt-data-task672-amazon-yelp- summarization Synthetic + au- dited 85.3K 0.00012%

  143. [151]

    schneiderkamplab/sapient-synth-flan- niv2-zsopt-data-task909-dialogre- prevalent-speakers Synthetic + au- dited 82.0K 0.00012%

  144. [152]

    schneiderkamplab/sapient-synth-flan- niv2-zsopt-data-task266-paper-reviews- reviewer-perspective Synthetic + au- dited 71.9K 0.0001%

  145. [153]

    schneiderkamplab/sapient-synth-flan- niv2-fsopt-data-task871-msmarco- question-generation Synthetic + au- dited 71.1K 0.0001%

  146. [154]

    schneiderkamplab/sapient-synth- flan-niv2-fsopt-data-task908-dialogre- identify-familial-relationships Synthetic + au- dited 68.4K 0.0001%

  147. [155]

    schneiderkamplab/sapient-synth-flan- niv2-zsopt-data-task906-dialogre- identify-names Synthetic + au- dited 57.5K 0.00008%

  148. [156]

    schneiderkamplab/sapient-synth-flan- niv2-zsopt-data-task264-paper-reviews- accept-reject Synthetic + au- dited 49.3K 0.00007%

  149. [157]

    schneiderkamplab/sapient-synth- flan-niv2-fsopt-data-task907-dialogre- identify-relationships Synthetic + au- dited 46.0K 0.00007%

  150. [158]

    schneiderkamplab/sapient-synth-flan- niv2-zsopt-data-task871-msmarco- question-generation Synthetic + au- dited 35.6K 0.00005%

  151. [159]

    schneiderkamplab/sapient-synth-flan- niv2-zsopt-data-task908-dialogre- identify-familial-relationships Synthetic + au- dited 17.3K 0.00002%

  152. [160]

    schneiderkamplab/sapient-synth-flan- niv2-zsopt-data-task907-dialogre- identify-relationships Synthetic + au- dited 13.4K 0.00002% 19 continued from previous page # Dataset Form Tokens/epoch Share

  153. [161]

    Table 11: Overview of benchmarks and their configuration

    schneiderkamplab/sapient-synth-flan- niv2-zsopt-data-task1370-newscomm- classification Synthetic + au- dited 1.9K<0.00001% B List of Evaluation Datasets Table 11 lists all considered evaluation datasets with their precise source as HuggingFace identi- fier, and the number of i...

Pith tools

Reviewed August 14, 2026 · model on record in the stance chip above.