wepiqx/MERNIK-rlmini-kld

MERNIK RL-mini KLD labels (the crown jewels)

Per-tensor damage on reasoning tasks, not wikitext — the label upgrade the whole Gnom program was waiting for. Each row: one unit dropped one tier, damage measured on rl_reasoning_mini (343 prompts, 5 RL-oss subsets, EN+ZH code/reasoning).

Why this file matters. PPL-damage labels saturate: a reasoning-critical layer and a filler layer can show the same wiki damage (the Perplexity Illusion). KLD against the unperturbed model is fidelity-sensitive, and the extra columns below are the first attempt to price how the distribution breaks, not just how much.

Columns (all damage_* = perturbed minus baseline):

Column	Meaning
damage	PPL damage (legacy anchor)
damage_kld	mean token-KLD vs reference (the allocator column)
damage_kld_max	worst-token KLD — tail detonations live here (up to 20.85 measured). The mechanism that breaks reasoning: the model generates 95% of a chain-of-thought fine, then on one critical token (say, > vs <) the distribution collapses — one token and the whole answer is garbage. Mean KLD cannot see it; max KLD can
damage_kld_median	median token-KLD — robust core damage. mean >> median means rare but catastrophic failures; mean ≈ median means uniform, predictable damage
damage_dp_rms / dp_rms	RMS of top-1 probability shift — decisiveness damage (both directions: lost confidence and artificial sharpening)
damage_same_top	fraction of positions keeping the argmax — rank before values. same_top = 0.80 means the model rewrites its decisions on every fifth token even when KLD looks fine; for code, one flipped token in if (x > y) breaks the program

Files.

• damage_17_rlmini.jsonl — Spark-1.7B (Q5_K base, 112 units + baseline), 113 rows, all 9 columns
• damage_4b_rlmini.jsonl — Spark-X2.5-4B (XHToken/Spark-X2.5-4B, Q5_K base), 145 rows (144 units, zero holes)

Censored units: "damage": null, "retryable": false — never impute.

Known nulls (read before training). The viewer shows null cells and they are honest, each with a reason:

• BASELINE rows: all kld*/damage_* are null — the baseline is the reference itself, there is nothing to compare it against.
• 4B rows without KLD columns: the 4B RL-mini leg measured PPL damage only (KLD needs the reference logits beside it; the 17B leg has both). damage is real, damage_kld* is null — absence of measurement, not zero damage.
• TERMINAL rows: every damage null with "retryable": false — the unit could not be measured at that tier at all. Drop, don't fill.

What's next. More of these: RL-mini-style labels are worth far more per row than wiki labels, so this repo keeps growing while the wiki pool stays frozen. Part of the MERNIK program; the full label archive (pools, legacy sweeps, failure ledgers) is wepiqx/MERNIK-gnom-labels.

---

If this work saved you a GPU-week: github.com/sponsors/wepiqx