adamtwiss/t91-binpacks-filtered

Filtered LC0 T91 training data (min-v2.v6)

Leela Chess Zero T91 self-play data, converted to Stockfish binpack format, minimised, and filtered with the Stockfish NNUE data-filtering tool — the processing step that has been applied to earlier LC0 runs but not to T90/T91.

5 months (Jan–May 2026) · 4.44 B positions · 3.08 B trainable · 12.6 GB

Files

file	positions	skip-marked	trainable	size	md5
test91-2026-01-jan-69gb6p.min-v2.v6.binpack	1,962,156,826	609,896,392	1,352,260,434	5.58 GB	755fbfcc6bb7b90a5e70f08b476e5fc4
test91-2026-02-feb-69gb6p.min-v2.v6.binpack	948,142,971	294,035,947	654,107,024	2.69 GB	83c08f803a640fcf5f9df2c2ecb67b01
test91-2026-03-mar-69gb6p.min-v2.v6.binpack	496,443,532	143,815,035	352,628,497	1.41 GB	1dee92ab4a1a01d780d1e4b3618c6f89
test91-2026-04-apr-69gb6p.min-v2.v6.binpack	422,847,598	122,263,996	300,583,602	1.20 GB	4c8c7734178d602f09b5f8378ed6abe5
test91-2026-05-may-69gb6p.min-v2.v6.binpack	609,640,539	189,633,553	420,006,986	1.74 GB	ceaffb8f79fb39aaef39bc1625dc6612
total	4,439,231,466	1,359,644,923	3,079,586,543	12.62 GB

Skip-marked positions carry score 32002 (VALUE_NONE). They are not deleted — trainers exclude them automatically (any max_score cap below 32002 does so; the Stockfish trainer defaults to 10000). "Trainable" is total − marked.

A note on file sizes

These files are considerably smaller than other T91 binpack datasets. That is an encoding difference, not a difference in content. Binpack can store a game either as full positions or delta-encoded from a stem position; minimize_binpack produces the latter.

Measured on April 2026:

dataset	bytes/position	positions	form
jshriver/t91-binpacks	26.18	449,683,241	full positions
xushawn/t91-binpacks-raw	18.19	433,470,115	full positions
this dataset	2.84	422,847,598	delta-encoded, filtered

Running minimize_binpack on a 982 MB slice of the jshriver file takes it from 26.12 to 2.26 bytes/position with no loss of positions (37,601,911 → 37,628,541; the tool re-derives chain boundaries). So all three sources hold substantially the same data — the published sizes differ by how it is stored. Position counts across the three differ by under 4%, most likely month-boundary choices.

Note that chain-stats-style measures do not reveal this: they report ply continuity (whether consecutive entries form a continuous game), which is similar across all three at ~105–115, and not whether the file physically delta-encodes those chains. Bytes-per-position is the measure that distinguishes them.

Source

Raw data from xushawn/t91-binpacks-raw, which packages LC0 T91 self-play output as Stockfish binpacks. Filenames preserve the source identifiers (69gb6p) so provenance is traceable.

Processing

Two steps, in this order, using a single pinned binary — the Stockfish fork at linrock/Stockfish, branch nnue-data-v7-3072, head d7e1fcd1b0 (build ARCH=x86-64-avx2):

transform minimize_binpack input_file <raw>  output_file <min>
transform filter_335a9b2d8a80 input_file <min> output_file <out>

1. minimize_binpack — repacks into proper chain-compressed form. The raw T91 packs ship in an expanded encoding (~18 bytes/position); minimised they are ~2.4. This matters beyond storage: trainers that weight files by sampled position density will mis-weight a 5× byte-heavy file in a mixed pool.
2. filter_335a9b2d8a80 — marks positions the NNUE trainer should skip: in-check, capture/promotion moves, the start position, and positions failing the depth-6 MultiPV-2 criteria. It only marks; it never deletes.

Resulting density is 2.83–2.85 bytes/position with average chain length 78–115, in line with existing filtered datasets (~2.6 B/position, chains ~105).

Sharding, position counting and encoding checks used Coda's split-binpack (splits at binpack block boundaries, so parts are independently valid and cat re-merges them), inspect-binpack and chain-stats.

Verification

Because filtering marks rather than deletes, output position count must equal input position count exactly. That was enforced per month, and every month matched. Filtering was distributed across 275 shards for January (fewer for smaller months); the collector refused to merge until every shard in the distribution manifest had a corresponding output.

Verify a download with:

md5sum -c MD5SUMS.txt

Distribution uniformity. Because shards are contiguous byte ranges in chronological order, a degraded run-period would appear as a contiguous block of anomalous shards. Sweeping per-shard label distributions across all 275 January and 132 February shards (~7 M positions each) found no such block: skip-mark rate stayed within 30.94–31.23%, and average |score| varied by under 2%, with no discontinuity anywhere in either series.

Placeholder zeros. LC0 data can contain positions scored 0 despite a decisively-scored predecessor and a decisive game result. Measured over a 200 M position sample, this dataset shows 0.00009%, against 0.00378% in a long-established filtered T80 file used as reference.

Does the filtering help?

Short answer: on T80 data yes, on this data no. In our testing this dataset trained a substantially weaker net than T80 of the same size, and the filtering made it worse rather than better. Please read this whole section before using it — the numbers are below.

A matched-control experiment on a separate month of T90 data (298,929,154 positions), training two NNUE nets with identical architecture, hyperparameters and schedule, differing only in whether the filter had run:

test	result
filtered vs raw (paired SPRT, 10+0.1)	+28.6 ± 13.0 Elo, LLR 3.01, H1

The filtered net won while training on 100.9 M fewer positions — the discarded third was net-negative for training.

...but it did not reproduce on this data — please read this before using it

Repeating the same paired design on this dataset gave the opposite answer, and the T91 corpus itself measured much weaker than T80. Each comparison was run twice, under two different training recipes:

comparison (matched trainable positions)	recipe A	recipe B
this data vs the same months unfiltered	−17.9 ± 9.1	−22.0 ± 10.0
this data vs an equal quantity of filtered T80	−64.2 ± 16.2	−99.6 ± 21.0

Both replicated with the same sign. Two conclusions, in order of importance:

1. T91 trained a much weaker net than T80 at equal size (~100 Elo). Most of the gap is the corpus, not the filtering.
2. The filtering made it worse, not better (~20 Elo). Note that trainers commonly discard in-check, capture/promotion and opening positions at load time already — if yours does, the only rule doing additional work is the MultiPV one, which removes quiet positions whose best continuation is tactical. That appears to be harmful here.

We would not recommend substituting this for known-good training data. These are small SB100 nets on one engine's recipe, so your setup may differ — but the results were consistent across recipes and we would rather publish them than leave the T80 result standing unqualified.

Although our testing found this data weaker (in the Coda engine) than T80 data, we are keeping it available in case it is useful for people to investigate and test with. We are fairly confident the processing and filtering itself is correct, and that the weaker nets we got are a property of the underlying data rather than of how we handled it.

Known characteristics

1. No time-based or run-phase exclusion was applied — filtering is purely per-position. See the uniformity check under Verification.

2. March and April were produced in the opposite order (filter, then minimise) before the ordering was settled. Consequence: their totals are lower (minimisation drops rows already marked) and average chain length is 78–82 rather than 113–115. Trainable position counts are unaffected — that invariant was verified for both — so training content is equivalent, but the files are not byte-for-byte comparable with the other three.

3. This is not all of T91. jshriver/t91-binpacks (unfiltered — we checked, it contains no skip-marked positions) additionally covers Nov 2025, Dec 2025 and June 2026, which the source used here does not include. For the overlapping months the two sources agree closely: for April, 433,470,115 positions here vs 449,683,241 there (96.3%), the remainder most likely a month-boundary difference.

4. The filter used is the current filter_335a9b2d8a80, which marks somewhat more than the older filters that produced earlier published .v6 datasets (~31% here vs ~18%). Most of the difference is in-check and capture positions, which trainers discard at load time regardless.

Credits

• Leela Chess Zero — the T91 run and the self-play data this is derived from.
• Stockfish — the binpack format and the data tooling; every processing step here is a Stockfish binary.
• linrock — the Stockfish fork carrying the minimize_binpack and filter_335a9b2d8a80 tools, and the earlier filtered datasets that established this pipeline.
• xu-shawn — for packaging T91 as binpacks in the first place, and for generous help and guidance throughout this work.

Released under ODbL, matching the source dataset.