repremover-xl

7,281 anti-repetition roleplay conversations. Each row is a real RP conversation containing an assistant turn that largely repeated an earlier turn, cut at that point, with the repeating turn replaced by a rewritten continuation. The intent: supervision located at the decision point where models tend to loop.

Built for spoomplesmaxx-thrasher-24B to reconstruct the idea behind the gated Dans-Prosemaxx-RepRemover-1. The full generation script (gen_repremover_xl.py) ships in this repo — point iter_rows() at your own conversation files and it rebuilds the dataset against your corpus.

⚠ How we recommend using it

Don't finetune on this lane alone. A single-lane topup on a narrow behavioral axis risks training a delta much broader than the axis you wanted (style drift, register washout). The pattern we use ourselves:

BLEND      25–35% of this dataset, 65–75% replay of your main SFT mix
           (weighted toward the register you care about)
LR         ~8× below your SFT peak (we anneal at ~1e-6 after an 8e-6 run)
LENGTH     one short epoch of the blend, not more
GATE       measure repetition behaviorally before AND after — with a
           multi-turn battery, not loss. Loss will look great either way.

We make no claims here about outcomes — run your own gate and believe your own numbers.

The multi-turn battery we gate with (battery_mt.py — loop-bait and stall-bait episodes, cross-turn Jaccard gates, 6+ repeats per episode) is published in the thrasher repo's eval/. Single-run battery numbers are noise; six repeats is the floor we trust.

Schema

{
  "id":       "repremover:<source conversation id>:<turn index>",
  "lane":     "repremover",
  "family":   "<source dataset shortname>",
  "source":   "repremover:<original hub dataset id>",
  "messages": [{"role": "system|user|assistant", "content": "..."}]
}

The final message is always the rewritten assistant turn. Everything before it is the original conversation, verbatim — including the repetition pressure that made the original model loop. That context is the training signal as much as the rewrite is.

How it was built

Mining. Across ~55K RP conversations (carded RP logs, think-stripped RP reasoning logs, and the PersonalityEngine RP family), find assistant turns whose word-set Jaccard similarity against any earlier assistant turn in the same conversation exceeds 0.45 — a turn that measurably re-treads. Up to 3 such sites per conversation (capped so heavy repeaters don't dominate). 20,126 sites qualified.

Rewriting. Truncate before the repeating turn; deepseek-4-pro (via the deepseek-chat API alias, temp 0.7) rewrites it under instructions to advance the scene, keep character and format, and reuse no distinctive phrasing. The prompt includes an explicit word-count target — this matters more than it looks (see stats).

Acceptance. A rewrite is kept only if it clears all four gates:

style     no Name:-prefix opener, no AI-break, no meta preamble
length    within [0.4×, 2.5×] of the turn it replaces
jaccard   < 0.35 against EVERY earlier assistant turn (stricter than mining)
beats     asterisk action beats iff the conversation uses them

The stats are a lesson in prompt specificity. Pass 1 asked for "similar length" in prose: 12% acceptance, with 86% of all rejections on the length gate — the rewriter simply ignored a vague instruction. Pass 2 added the numeric word-count target and re-ran every rejected site: acceptance jumped to 29%, length rejections collapsed, and the Jaccard gate became the main filter (which is what you want — it means the gates are now rejecting echoes, not formatting accidents).

            pass 1      pass 2 (word-count target)
accepted    2,491       5,086
rej:length  14,643      6,354
rej:jaccard 191         4,790
rej:beats   390         1,173
rej:style   154         185

Final: 7,281 unique conversations (~94M chars): 89% beat-free dialogue registers, 11% asterisk-beat prose. If your target style uses action beats, consider rebalancing before blending — the supervised style mix is a lesson of its own.

Provenance

Source conversations: anthracite c2 logs, Gryphe Aesir charcards, PJMixers log collections, bluemoon, chub/aicg logs, and the PocketDoc PersonalityEngine RP family (Personamaxx-VN, Failuremaxx, Prosemaxx Adventure/Gutenberg, Kinomaxx). Each row's source field names its origin. Rewritten turns generated by deepseek-4-pro, 2026-08. Adult content is present, in line with the source RP corpora — this is a dataset for roleplay models, built by adults, for adults.

---

mimids tooling · built for thrasher (mimids 02) · Apache 2.0