iamPi/russet_shrike

Duel-cover delta, epoch 18 → 20

The rows that changed between the epoch-18 and epoch-20 builds of the 60% duel cover published at iamPi/cobalt_heron.

Turn ids only — no prefixes, thoughts, or actions. Fetch the text from https://data.affine.io using the reconstruction loop in the parent dataset's README (note: that host returns 403 without a User-Agent header).

	rows
epoch-18 cover	56,603
epoch-20 cover	67,084
overlap	55,293
added (delta_e18_to_e20_added.jsonl)	11,791
removed (delta_e18_to_e20_removed.jsonl)	1,310

97.69% of the epoch-18 cover survives; 82.42% of the epoch-20 cover was already held. If you have already materialized the epoch-18 rows, applying this delta gets you to the epoch-20 cover without re-fetching 55,293 unchanged turns.

Files

delta_e18_to_e20_added.jsonl — one object per new turn, same schema as the parent dataset:

{"turn_id":"…:3","chunk_key":"views/duel_turns@v4/chunks/view_0020_0000.jsonl.gz","traj_line":18,"node_id":8,"action_kind":"text"}

delta_e18_to_e20_removed.jsonl — turn ids to drop, with the reason:

{"turn_id":"…:7","reason":"stratum_grew_past_cutoff"}

delta_e18_to_e20_stats.json — counts, sha256s, and the breakdowns below.

What was added

11,791 rows / 574 MB of prefix, worth 108.7 of the 1,300 turns in a duel slice. These come from strata that are new at epoch 20 or that the size-ranked rule newly admits.

source	rows		dialect	rows		phase	rows
swesmith	4,379		bash	5,839		early	4,471
terminal_lego	1,840		tool_call	4,594		late	3,900
scaleswe	1,604		text	1,358		mid	3,420
multiswe	1,272
swerebench_v2	1,200
swelego	563
affine_wiki	493
affine_agent	262
terminal_bench_2	80
nl2repobench	54
affine_nl2lib	44

Every text-dialect row in the cover is in this delta. text was admitted to allowed_action_kinds at weight_version_key 13 (2026-09-09), so no epoch-18 build contains any. Those 1,358 rows need special handling: the text action is the entire visible reply, so the thought parsed out of duel_turns@v4 is always empty (median and max length 0 across 300 sampled turns). Source their thoughts from traces/, which retains reasoning_content, or drop them — training them as-is teaches an empty thought and fails the min_thought_chars = 80 crown gate.

What was removed

1,310 rows, worth only 1.9 duel turns. None of them left the corpus — all 1,310 still exist at epoch 20. They fell out because their strata grew past the ≤28 turn cutoff:

source	rows
affine_agent	1,289
swelego	21

affine_agent grew from 2,183 to 51,457 turns between epochs 18 and 20 while gaining only 336 strata — its median trajectory is now 97 turns, so a rule that ranks strata by size drops nearly all of it.

Provenance

Corpus epoch 20, view duel_turns@v4, manifest 61f92695a5aaab0ce5148a362a8a8cc3cea59d269145335bb2f1e5918eebacbb (283,828 turns / 9,903 strata). Both sides were produced by build_cover.py in the parent dataset, run against the epoch-18 and epoch-20 index shards.