Suggestions are the perishable half. Every provider answers eth_gasPrice and eth_maxPriorityFeePerGas for right now, computed by that node from its own mempool view, and none of them publishes a history of its own answers. No archive holds a historical mempool, so a suggestion made last month exists only if somebody recorded it at the time. What a block actually required, by contrast, is on chain forever. That asymmetry is why this join can be published at all: anyone can recompute the outcome, and nobody can go back and find out what was advised.

name	one row is
`e29_gas_estimators`	one provider's gas suggestion at one moment, tagged with the block it was standing on
`e29_eth_block_fees`	one block: base fee, priority-fee percentiles actually paid, gas used ratio
`e30_gas_estimator_accuracy`	one provider's suggestion scored against the block it was advice about

The providers disagree, and the disagreement is not latency

This was the first thing checked, because it is the way the dataset could have been worthless. An early pass showed a 6.3% spread while two of four providers sat one block behind, which would have made this a latency measurement wearing a disagreement costume. Comparing providers strictly within the same block, over six consecutive blocks:

quantity	median within-block spread
`eth_gasPrice`	7.3% (max 9.4%)
`eth_maxPriorityFeePerGas`	100%, on every block

It is systematic rather than noisy. One provider returned a zero priority fee on every block sampled and another on most, while the rest agreed on a real value. Two of the disagreeing endpoints are MEV-protection RPCs that see genuinely different order flow, so this is a difference in what each node can see, not sampling error.

Reading the accuracy table

sufficient_priority answers the only question a wallet actually asks: would paying that suggestion have got the transaction into a block within the horizon. overpay_priority is the suggestion divided by what cleared, so 1.0 is exact and 3.0 means paying triple.

Matching needs no timestamps, which removes the usual way this kind of join goes wrong. Every suggestion carries the head block the provider reported at the moment it answered, so a suggestion made at head N is by construction advice about block N+1, and it is scored against blocks N+1 through N+3. Matching on wall-clock time would have to guess at propagation and sampling lag.

What counts as the rate that cleared

cleared_p10_gwei is the 10th percentile of priority fees actually paid by transactions in the block, from eth_feeHistory. Not the minimum: blocks routinely include zero-priority transactions placed by builders or delivered privately, and those measure a private arrangement rather than the price a stranger would have had to pay. The 5th percentile and the median travel alongside so you can pick a different definition rather than accept ours.

A cleared rate of exactly zero is common and genuine. It means the block accepted at least a tenth of its transactions at no tip at all, so sufficiency is trivially true and the overpay ratio is undefined rather than infinite. Those rows carry a null, never a large number.

Before you build on this

Ethereum mainnet only, and that is a measured decision rather than an omission. The L2s were probed and show no cross-provider spread worth recording: Base 0.0%, Optimism 0.0%, Polygon 0.0%, Arbitrum 0.3%. One sequencer posts one number and every node echoes it.
Seven providers, and they are the ones that answer without an API key. That is a selection, and it skews toward endpoints that serve retail wallets rather than institutional order flow.
A provider that fails is recorded with an error, never as a zero suggestion. A failed read and a genuine zero tip are different facts and the panel keeps them apart.
The spread is measured in percent, and Ethereum has been quiet during early collection. A 7% spread on a 0.5 gwei base is a few hundred wei. Whether divergence widens under congestion, when mempool views differ most, is exactly what a long panel answers and a short one cannot.
Suggestions are sampled every 20 seconds against roughly 12 second blocks, so most blocks are covered but not all. head_block is on every row, so coverage is checkable rather than assumed.
e29_eth_block_fees is the freely available half. It is carried so the join is self contained, not because it is scarce.

Partitions are parquet, one file per collection window, under dataset/YYYY/MM/. Every dataset here carries a FIXED 7-day sample WINDOW starting at its own first day of collection, together spanning 2026-08-31 to 2026-09-06, so you can check schema, coverage and quality before asking for more. It does not advance, so there is nothing to gain by re-downloading it. The full history is held privately, available on request.

from huggingface_hub import snapshot_download
import pandas as pd, glob

path = snapshot_download("dataforge-labs/ethereum-gas-estimator-accuracy", repo_type="dataset",
                         allow_patterns="e30_gas_estimator_accuracy/**")
df = pd.concat(map(pd.read_parquet,
                   glob.glob(f"{path}/e30_gas_estimator_accuracy/**/*.parquet", recursive=True)))

Coverage

e0_run_manifest lists every collection window with its poll counts and failure counts, and is published in full rather than windowed. Gaps between windows are real, cannot be filled in afterwards, and nothing here is interpolated.

License and contact

ODC-BY: use it freely, credit "DataForge (dataforge-labs)". Questions and requests for the full history via the discussions tab.

Downloads last month: 60

Total file size:

5.5 MB

Ethereum gas estimator accuracy