Back to .md Directory

RWKV-Evals

RWKV Evaluation Data presents RWKV's performance across various large language model benchmarks, including Uncheatable Eval, MMLU, RULER, and LongBench.

May 2, 2026
0 downloads
6 views
ai rag eval
View source

RWKV-Evals

RWKV Evaluation Data presents RWKV's performance across various large language model benchmarks, including Uncheatable Eval, MMLU, RULER, and LongBench.

Uncheatable Eval Test

::: tip Uncheatable Eval is an "uncheatable evaluation" that uses real-time data, such as the latest academic papers and news articles, to assess the true modeling capabilities and generalization abilities of open-source large language models.

The result of the Uncheatable Eval test is the compression ratio; therefore, a lower score indicates better model performance. :::

Below is a comparison of Uncheatable Eval scores between RWKV and other models:

14B Parameter Models

NameParams (B)Average (lower=better)ao3 englishbbc newswikipedia englisharxiv computer sciencearxiv physicsgithub cppgithub python
RWKV7-g0b-13.3b13.2696.8439.8488.2027.6367.1087.3804.0263.892
Qwen3-14B-Base14.7686.84510.5698.4457.9427.0017.2103.4393.312
gemma-3-12b-pt12.1876.94510.5407.9147.6077.2867.3873.8833.997
Qwen2.5-14B14.7706.95110.5588.3177.9447.2247.3923.6253.599
Mistral-Nemo-Base-240712.2486.97010.1658.1187.6427.2877.4554.0794.042
Motif-2-12.7B-Base12.7047.09910.6288.3287.8977.1347.4044.1894.114
Llama-2-13b-hf13.0167.54010.6558.3077.9017.9938.1224.7955.009

7B Parameter Models

NameParams (B)Average (lower=better)ao3 englishbbc newswikipedia englisharxiv computer sciencearxiv physicsgithub cppgithub python
Qwen3-8B-Base8.1917.09110.8908.7188.2557.2077.4653.6173.482
Meta-Llama-3-8B8.0307.16210.6198.2957.7857.5367.5414.1744.181
RWKV7-g0a3-7.2b-20251029-ctx81927.1997.22210.1648.4807.9967.4407.7474.3784.347
Qwen2.5-7B7.6167.32311.0798.7298.4497.5397.7923.8683.806
Falcon-H1-7B-Base7.5867.33910.9588.5768.2257.4037.5694.2514.392
Mistral-7B-v0.17.2427.40610.6628.3067.9767.7457.9034.6124.635
Hunyuan-7B-Pretrain7.5057.54111.5098.9878.4997.6538.1084.2013.829
falcon-mamba-7b7.2737.54810.7608.9588.5897.6747.7374.4374.680
Zamba2-7B7.3577.58210.7028.6278.0747.8438.1244.8334.869
Minitron-8B-Base8.2727.58210.8358.6548.2847.8568.2304.5084.708
Olmo-3-1025-7B7.2987.59511.1018.7848.5227.4907.9474.9304.394
RWKV-x060-World-7B-v3-20241112-ctx40967.6367.63310.6298.7538.2887.9368.1094.7864.929

3B Parameter Models

NameParams (B)Average (lower=better)ao3 englishbbc newswikipedia englisharxiv computer sciencearxiv physicsgithub cppgithub python
RWKV7-g1a4-2.9b-20251118-ctx81922.9487.48610.4818.8008.3107.7128.0724.5534.474
Llama-3.2-3B3.2137.64311.2198.7018.3657.9288.0654.6614.562
Qwen2.5-3B3.0867.72211.5759.1398.8957.9118.2204.2034.113
SmolLM3-3B-Base3.0757.78411.1878.9058.6118.0978.6314.5134.546
RWKV-x070-World-2.9B-v3-20250211-ctx40962.9487.80010.8128.9098.5018.0498.3074.9555.066
stablelm-3b-4e1t2.7957.90711.2118.8158.4348.2998.4764.9065.207
Falcon-H1-3B-Base3.1497.93611.6859.1588.9107.8918.1614.8324.917
recurrentgemma-2b2.6838.05211.6328.9518.8358.4018.4884.8975.157
RWKV-x060-World-3B-v2.1-20240417-ctx40963.1008.14711.0059.1618.8158.4518.5595.4795.561
mamba2attn-2.7b2.6988.20411.4369.2468.9478.4748.2365.3365.751

1.5B Parameter Models

NameParams (B)Average (lower=better)ao3 englishbbc newswikipedia englisharxiv computer sciencearxiv physicsgithub cppgithub python
Qwen3-1.7B-Base1.7217.96512.0169.7439.3527.9368.3504.2604.095
RWKV7-g1b-1.5b-20251015-ctx81921.5277.96910.9729.2508.8438.1108.5375.0415.027
Qwen2.5-1.5B1.5448.12412.1149.5629.3938.2708.6464.5024.384
RWKV-x070-World-1.5B-v3-20250127-ctx40961.5278.23111.2739.3208.9658.4318.7585.3855.483
SmolLM2-1.7B1.7118.29811.5369.3739.3518.5479.0475.0805.152
Llama-3.2-1B1.2368.30612.0369.3319.0978.5568.7555.2675.101
Index-1.9B2.1738.34011.8319.4939.0698.4978.5615.3805.547
stablelm-2-1_6b1.6458.39611.7619.2378.9438.7629.0885.5585.425
Falcon-H1-1.5B-Deep-Base1.5558.50512.1449.6669.4828.4078.9685.4975.368
RWKV-x060-World-1B6-v2.1-20240328-ctx40961.6008.56411.4349.5559.2768.8228.9905.9065.968
Falcon-H1-1.5B-Base1.5558.63912.2879.7969.6458.5079.0895.6355.514
mamba2-1.3b1.3448.69911.9449.7109.4638.9258.7145.8516.286
RWKV-5-World-1B5-v2-20231025-ctx40961.5788.71511.5959.7319.4518.9779.1036.0396.110
mamba-1.4b-hf1.3728.80612.0269.7839.5529.0818.8365.9586.408

MMLU Test

::: tip MMLU (Massive Multitask Language Understanding) is a widely recognized benchmark of general knowledge attained by AI models. It covers a broad range of topics from 57 different categories, covering elementary-level knowledge up to advanced professional subjects like law, physics, history, and computer science. :::

ModelMMLUMMLU COT
RWKV7-g0b-13.3b0.7650.827
RWKV7-g0a3-7.2b0.6510.723
RWKV7-g1a4-2.9b0.6130.675
RWKV7-g1b-1.5b0.5050.542

MMLU Pro Test

::: tip MMLU-Pro dataset is a more robust and challenging massive multi-task understanding dataset tailored to more rigorously benchmark large language models' capabilities. This dataset contains 12K complex questions across various disciplines. :::

ModelMMLU-PROMMLU-PRO COT
RWKV7-g0b-13.3b0.5020.612
RWKV7-g0a3-7.2b0.3590.521
RWKV7-g1a4-2.9b0.3240.43
RWKV7-g1b-1.5b0.2220.292

GSM8K Test

::: tip GSM8K (Grade School Math 8K) is a dataset of 8.5K high quality linguistically diverse grade school math word problems. The dataset was created to support the task of question answering on basic mathematical problems that require multi-step reasoning. :::

ModelGSM8K
RWKV7-g0b-13.3b0.923
RWKV7-g0a3-7.2b0.839
RWKV7-g1a4-2.9b0.773
RWKV7-g1b-1.5b0.585

MATH500 Test

::: tip MATH-500 is an authoritative benchmark for measuring the mathematical reasoning capabilities of AI models. It contains 500 challenging mathematical problems covering algebra, geometry, calculus, probability, statistics, and other fields. :::

ModelMATH500
RWKV7-g0b-13.3b0.768
RWKV7-g0a3-7.2b0.678
RWKV7-g1a4-2.9b0.482
RWKV7-g1b-1.5b0.298

IFEval

::: tip IFEval (Instruction-Following Evaluation) is a straightforward and easy-to-reproduce evaluation benchmark. It focuses on a set of "verifiable instructions" such as "write in more than 400 words" and "mention the keyword of AI at least 3 times". :::

ModelIFEval (strict prompt-level)
RWKV7-g0b-13.3b0.689
RWKV7-g0a3-7.2b0.58
RWKV7-g1a3-2.9b0.51
RWKV7-g1b-1.5b0.421

RULER Test

::: tip RULER is a new LLM testing method that optimizes and extends the NIAH (Needle In A Haystack) test. It includes four types of tasks: Enhanced Retrieval (Extended NIAH), Multi-hop Tracing, Information Aggregation (CWE, FWE), and QA with interference. :::

Enhanced Needle In A Haystack (NIAH)

RULER includes the Enhanced Needle In A Haystack (NIAH) test, divided into four sub-tasks to evaluate the model's retrieval capabilities:

Sub-taskBrief Description
Single NIAH (S-NIAH)Tests the model's ability to handle a single input and a single target output.
Multi-keys NIAH (MK-NIAH)Tests the model's ability to handle multiple key-value pairs, where each key is associated with a single output.
Multi-values NIAH (MV-NIAH)Tests the model's ability to handle multiple key-value pairs, where each key is associated with multiple values or outputs.
Multi-queries NIAH (MQ-NIAH)Tests the model's ability to synthesize and generate corresponding results under multiple query conditions.

Single NIAH test results are as follows:

ModelNIAH_single_1NIAH_single_2NIAH_single_3
RWKV-6-7B-v2.1-4k10098.6795
Llama2-7B-4k97.610096.8
Mamba-2.8B-4k100191
Mamba-1.4B-4k94215
RWKV-6-3B-v2.1-4k1008879
RWKV-6-1.6B-v2.1-4k985355

NIAH-Multi-keys test results are as follows:

ModelNIAH_multikey_1NIAH_multikey_2NIAH_multikey_3
RWKV-6-7B-v2.1-4k48.337.671.33
Llama2-7B-4k10084.460
Mamba-2.8B-4k701
Mamba-1.4B-4k800
RWKV-6-3B-v2.1-4k3610
RWKV-6-1.6B-v2.1-4k2510

Multi-values and Multi-queries NIAH test results are as follows:

ModelNIAH_multivalueNIAH_multiquery
RWKV-6-7B-v2.1-4k80.4283.67
Llama2-7B-4k9496.7
Mamba-2.8B-4k0.751.25
Mamba-1.4B-4k5.254.75
RWKV-6-3B-v2.1-4k38.540.75
RWKV-6-1.6B-v2.1-4k2520.75

Variable Tracking (VT)

::: tip Multi-hop Tracing - Variable Tracking: This task mainly checks whether the model can successfully identify and track entities (variables) and reference relationships with multi-hop connections within a long context. For example, given the assignment $X_1 = V$, then $X_2 = X_1$, $X_3 = X_2$ ... finally returning all variable names pointing to the value $V$. :::

ModelMulti-hop Tracing
RWKV-6-7B-v2.1-4k7.53
Llama2-7B-4k63.12
Mamba-2.8B-4k45
Mamba-1.4B-4k23.4
RWKV-6-3B-v2.1-4k11.8
RWKV-6-1.6B-v2.1-4k1.4

Information Aggregation (CWE, FWE)

::: tip Information Aggregation (CWE, FWE): This task involves Common Words Extraction (CWE) and Frequent Words Extraction (FWE), used to test the model's ability to aggregate common information across long contexts. :::

ModelCommon Words Extraction (CWE)Frequent Words Extraction (FWE)
RWKV-6-7B-v2.1-4k38.678.33
Llama2-7B-4k73.0478.8
Mamba-2.8B-4k253
Mamba-1.4B-4k15.557.33
RWKV-6-3B-v2.1-4k30.362.67
RWKV-6-1.6B-v2.1-4k1146.33

Question Answering (QA)

::: tip Question Answering (QA): This task adds distracting information to the input of existing short-context QA datasets to evaluate QA capabilities under various context sizes. :::

Modelqa_1qa_2
RWKV-6-7B-v2.1-4k4537
Llama2-7B-4k59.242
Mamba-2.8B-4k2318
Mamba-1.4B-4k2423
RWKV-6-3B-v2.1-4k3525
RWKV-6-1.6B-v2.1-4k3528

Data Source

RULER Data Source: https://github.com/Ojiyumm/RULER_RWKV

LongBench Test

::: tip LongBench is a benchmark for evaluating the long-text understanding capabilities of large language models.

LongBench consists of six major categories and twenty-one different bilingual (Chinese-English) tasks, covering critical long-text application scenarios such as Single-Document QA, Multi-Document QA, Summarization, Few-shot Learning, Synthetic Tasks, and Code Completion. :::

Below is a comparison of LongBench scores for RWKV and other models, with data tables presented by the six categories:

Single-Document QA

Single-Document QA includes the following four test tasks:

TaskTask Description
NarrativeQAQA based on stories or scripts, including understanding of characters, plots, themes, etc.
QasperQA based on single papers; questions are asked by NLP readers and answered by NLP practitioners.
MultiFieldQA-enAnswering English questions based on a single document from relatively diverse fields.
MultiFieldQA-zhAnswering Chinese questions based on a single document from relatively diverse fields.

Single-Document QA Test Results:

ModelNarrativeQAQasperMultiFieldQA-enMultiFieldQA-zh
GPT-3.5-Turbo-16k23.643.352.361.2
Llama2-7B-chat-4k18.719.236.811.9
LongChat-v1.5-7B-32k16.927.741.429.1
XGen-7B-8k18.018.137.714.8
InternLM-7B-8k12.116.723.433.6
ChatGLM2-6B-32k21.131.546.251.6
Vicuna-v1.5-7B-16k19.426.138.543.0
ChatGLM3-6B-32k26.043.351.762.3
Mamba_1B42.234.4411.3313.03
Mamba_2B82.324.898.156.83
Llama2-7B18.719.211.9036.8
Mistral-7B12.798.930.5517.91
RWKV-6-World-1B6-v2.14.5319.7922.9918.57
RWKV-6-World-3B-v2.12.8714.218.7821.49
RWKV-6-World-7b-v2.1-4k20.7540.236.0150.19

Multi-Document QA

Multi-Document QA includes the following four test tasks:

TaskTask Description
HotpotQAAnswering questions based on HotpotQA documents; involves many 2-hop questions written by native speakers based on two related paragraphs.
2WikiMultihopQAAnswering questions based on 2WikiMultihopQA data; composed of up to 5-hop questions synthesized via manually designed templates.
MuSiQueAnswering questions based on MuSiQue data; composed of simple questions requiring up to 4-hop reasoning.
DuReaderAnswering questions based on the Chinese DuReader dataset, containing 200k questions and 1M documents from Baidu Search and Baidu Zhidao.

Multi-Document QA Test Results:

ModelHotpotQA2WikiMQAMusiqueDuReader (zh)
GPT-3.5-Turbo-16k51.637.726.928.7
Llama2-7B-chat-4k25.432.89.45.2
LongChat-v1.5-7B-32k31.520.69.719.5
XGen-7B-8k29.721.110.311.0
InternLM-7B-8k28.722.89.011.1
ChatGLM2-6B-32k45.134.021.937.6
Vicuna-v1.5-7B-16k25.320.89.819.3
ChatGLM3-6B-32k54.444.940.444.78
Mamba_1B45.738.773.311.95
Mamba_2B85.498.453.4513.96
Llama2-7B25.432.89.45.2
Mistral-7B9.3911.174.5811.68
RWKV-6-World-1B6-v2.18.7211.863.9611.40
RWKV-6-World-3B-v2.16.799.644.1317.41
RWKV-6-World-7b-v2.1-4k22.7416.310.4928.01

Summarization

The Summarization category involves the following four test tasks:

TaskTask Description
GovReportSummarization task requiring summaries of government work reports.
QMSumSummarization task requiring summaries of meeting minutes based on user queries.
MultiNewsMulti-document summarization task requiring summaries based on multiple news articles.
VCSUMSummarization task requiring summaries of Chinese meeting minutes.

Summarization Test Results:

ModelGovReportQMSumMultiNewsVCSUM (zh)
GPT-3.5-Turbo-16k29.523.426.716.0
Llama2-7B-chat-4k27.320.825.80.2
LongChat-v1.5-7B-32k30.822.726.49.9
XGen-7B-8k27.320.526.22.2
InternLM-7B-8k9.715.922.812.4
ChatGLM2-6B-32k32.424.026.516.2
Vicuna-v1.5-7B-16k27.922.827.215.1
ChatGLM3-6B-32k36.823.927.917.8
Mamba_1B49.3410.8515.866.33
Mamba_2B810.4111.4218.946.1
Llama2-7B27.320.825.80.2
Mistral-7B28.8420.3222.794.1
RWKV-6-World-1B6-v2.117.5120.3621.5210.71
RWKV-6-World-3B-v2.119.212121.7610.18
RWKV-6-World-7b-v2.1-4k31.6421.3126.0615.19

Few-shot Learning

Few-shot Learning includes the following four test tasks:

TaskTask Description
TRECClassification task requiring question classification, containing 50 categories in total.
TriviaQASingle-document QA task, providing several Few-shot examples.
SAMSumDialogue summarization task, providing several Few-shot examples.
LSHTChinese classification task requiring news classification, containing 24 categories in total.

Few-shot Learning Test Results:

ModelTRECTriviaQASAMSumLSHT (zh)
GPT-3.5-Turbo-16k68.091.441.729.2
Llama2-7B-chat-4k61.577.840.719.8
LongChat-v1.5-7B-32k63.582.334.223.2
XGen-7B-8k65.577.825.320.5
InternLM-7B-8k52.077.821.215.2
ChatGLM2-6B-32k62.578.736.327.7
Vicuna-v1.5-7B-16k71.586.240.828.8
ChatGLM3-6B-32k79.087.138.242.0
Mamba_1B445.537.3312.568.5
Mamba_2B821.534.629.35
Llama2-7B61.577.840.719.8
Mistral-7B70.089.2643.7425.5
RWKV-6-World-1B6-v2.139.547.6413.5818.8
RWKV-6-World-3B-v2.151.557.1517.9515.2
RWKV-6-World-7b-v2.1-4k55.586.8944.2530.2

Synthetic Tasks

Synthetic Tasks include the following three test tasks:

TaskTask Description
PassageCountDetermine the total number of unique paragraphs among the given paragraphs.
PassageRetrieval-enGiven 30 English Wikipedia paragraphs, determine which paragraph the given summary belongs to.
PassageRetrieval-zhGiven several Chinese paragraphs from the C4 dataset, determine which paragraph the given summary belongs to.

Synthetic Tasks Test Results:

ModelPassage CountPassageRetrieval-enPassageRetrieval-zh
GPT-3.5-Turbo-16k4.571.077.5
Llama2-7B-chat-4k2.19.80.5
LongChat-v1.5-7B-32k1.030.57.6
XGen-7B-8k2.18.53.5
InternLM-7B-8k3.06.00.9
ChatGLM2-6B-32k1.577.064.5
Vicuna-v1.5-7B-16k6.54.55.0
ChatGLM3-6B-32k2.099.094.0
Mamba_1B40.453.323.81
Mamba_2B80.741.833.37
Llama2-7B2.19.80.5
Mistral-7B1.0512.516.75
RWKV-6-World-1B6-v2.104.254.16
RWKV-6-World-3B-v2.103.834.12
RWKV-6-World-7b-v2.1-4k534.554.22

Code Completion

Code Completion includes the following two test tasks:

TaskTask Description
LCCGiven a long piece of code, predict the next line of code.
RepoBench-PGiven code from multiple files in a GitHub repository (including inter-file dependencies), predict the next line of code.

Code Completion Test Results:

ModelLCCRepoBench-P
GPT-3.5-Turbo-16k54.753.6
Llama2-7B-chat-4k52.443.8
LongChat-v1.5-7B-32k53.055.3
XGen-7B-8k38.638.6
InternLM-7B-8k44.128.8
ChatGLM2-6B-32k55.649.9
Vicuna-v1.5-7B-16k51.043.5
ChatGLM3-6B-32k57.6654.76
Mamba_1B444.3341.86
Mamba_2B839.5324.38
Llama2-7B52.443.8
Mistral-7B70.6459.7
RWKV-6-World-1B6-v2.139.540.44
RWKV-6-World-3B-v2.140.0141.35
RWKV-6-World-7b-v2.1-4k73.8454.1

Comprehensive Score Comparison of RWKV, Mamba, and Llama2

ModelSingle DocQFew-shcSummarizationMulti Doc QACode CompletionSyntetic
RWKV-6-World-1B6-v2.116.47029.86817.5258.98539.9702.803
RWKV-6-World-3B-v2.114.33535.44318.0389.49340.6802.650
RWKV-6-World-7b-v2.1-4k36.78854.20323.55019.38563.97031.240
Mamba_1B47.75825.97310.5957.43843.0952.527
Mamba_2B85.54817.60511.7187.83831.9551.980
Llama2-7B21.65049.95018.52518.20048.1004.133
Mistral-7B17.53852.83319.0139.20565.1710.100

Data Source

Evaluation Data Source: https://github.com/Ojiyumm/LongBench_RWKV

Related Documents