Datasets:

wepiqx
/

MERNIK-gnom-labels

Error code:   DatasetGenerationError
Exception:    CastError
Message:      Couldn't cast
unit: string
ppl: double
kld: double
kld_max: double
kld_median: double
dp_rms: double
same_top: double
tensors: list<item: string>
  child 0, item: string
damage: double
damage_kld: double
damage_kld_max: double
damage_kld_median: double
damage_dp_rms: double
damage_same_top: double
to
{'unit': Value('string'), 'ppl': Value('float64'), 'kld': Value('float64'), 'tensors': List(Value('string')), 'damage': Value('float64'), 'damage_kld': Value('float64')}
because column names don't match
Traceback:    Traceback (most recent call last):
                File "/usr/local/lib/python3.14/site-packages/datasets/builder.py", line 1827, in _prepare_split_single
                  for key, table in generator:
                                    ^^^^^^^^^
                File "/src/services/worker/src/worker/job_runners/config/parquet_and_info.py", line 613, in wrapped
                  for item in generator(*args, **kwargs):
                              ~~~~~~~~~^^^^^^^^^^^^^^^^^
                File "/usr/local/lib/python3.14/site-packages/datasets/packaged_modules/json/json.py", line 343, in _generate_tables
                  self._cast_table(pa_table, json_field_paths=json_field_paths),
                  ~~~~~~~~~~~~~~~~^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
                File "/usr/local/lib/python3.14/site-packages/datasets/packaged_modules/json/json.py", line 132, in _cast_table
                  pa_table = table_cast(pa_table, self.info.features.arrow_schema)
                File "/usr/local/lib/python3.14/site-packages/datasets/table.py", line 2378, in table_cast
                  return cast_table_to_schema(table, schema)
                File "/usr/local/lib/python3.14/site-packages/datasets/table.py", line 2306, in cast_table_to_schema
                  raise CastError(
                  ...<3 lines>...
                  )
              datasets.table.CastError: Couldn't cast
              unit: string
              ppl: double
              kld: double
              kld_max: double
              kld_median: double
              dp_rms: double
              same_top: double
              tensors: list<item: string>
                child 0, item: string
              damage: double
              damage_kld: double
              damage_kld_max: double
              damage_kld_median: double
              damage_dp_rms: double
              damage_same_top: double
              to
              {'unit': Value('string'), 'ppl': Value('float64'), 'kld': Value('float64'), 'tensors': List(Value('string')), 'damage': Value('float64'), 'damage_kld': Value('float64')}
              because column names don't match
              
              The above exception was the direct cause of the following exception:
              
              Traceback (most recent call last):
                File "/src/services/worker/src/worker/job_runners/config/parquet_and_info.py", line 1369, in compute_config_parquet_and_info_response
                  parquet_operations, partial, estimated_dataset_info = stream_convert_to_parquet(
                                                                        ~~~~~~~~~~~~~~~~~~~~~~~~~^
                      builder, max_dataset_size_bytes=max_dataset_size_bytes
                      ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
                  )
                  ^
                File "/src/services/worker/src/worker/job_runners/config/parquet_and_info.py", line 948, in stream_convert_to_parquet
                  builder._prepare_split(split_generator=splits_generators[split], file_format="parquet")
                  ~~~~~~~~~~~~~~~~~~~~~~^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
                File "/usr/local/lib/python3.14/site-packages/datasets/builder.py", line 1694, in _prepare_split
                  for job_id, done, content in self._prepare_split_single(
                                               ~~~~~~~~~~~~~~~~~~~~~~~~~~^
                      gen_kwargs=gen_kwargs, job_id=job_id, **_prepare_split_args
                      ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
                  ):
                  ^
                File "/usr/local/lib/python3.14/site-packages/datasets/builder.py", line 1879, in _prepare_split_single
                  raise DatasetGenerationError("An error occurred while generating the dataset") from e
              datasets.exceptions.DatasetGenerationError: An error occurred while generating the dataset

Need help to make the dataset viewer work? Make sure to review how to configure the dataset viewer, and open a discussion for direct support.

unit string	ppl float64	kld float64	tensors list	damage float64	damage_kld float64
BASELINE	59.6195	0.165629	null	null	null
attn_gate@0+attn_qkv@0	60.282	0.169927	[ "blk.0.attn_gate.weight", "blk.0.attn_qkv.weight" ]	0.6625	0.004298
attn_output@0	58.8337	0.172283	[ "blk.0.attn_output.weight" ]	-0.7858	0.006654
ffn_down@0	60.0583	0.163231	[ "blk.0.ffn_down.weight" ]	0.4388	-0.002398
ffn_gate@0+ffn_up@0	57.3043	0.170214	[ "blk.0.ffn_gate.weight", "blk.0.ffn_up.weight" ]	-2.3152	0.004585
attn_gate@1+attn_qkv@1	60.3979	0.171607	[ "blk.1.attn_gate.weight", "blk.1.attn_qkv.weight" ]	0.7784	0.005978
attn_output@1	60.0129	0.166644	[ "blk.1.attn_output.weight" ]	0.3934	0.001015
ffn_down@1	60.5941	0.167464	[ "blk.1.ffn_down.weight" ]	0.9746	0.001835
ffn_gate@1+ffn_up@1	58.8503	0.171526	[ "blk.1.ffn_gate.weight", "blk.1.ffn_up.weight" ]	-0.7692	0.005897
attn_gate@10+attn_qkv@10	60.1763	0.169	[ "blk.10.attn_gate.weight", "blk.10.attn_qkv.weight" ]	0.5568	0.003371
attn_output@10	61.0118	0.169157	[ "blk.10.attn_output.weight" ]	1.3923	0.003528
ffn_down@10	59.627	0.168234	[ "blk.10.ffn_down.weight" ]	0.0075	0.002605
ffn_gate@10+ffn_up@10	58.7875	0.172625	[ "blk.10.ffn_gate.weight", "blk.10.ffn_up.weight" ]	-0.832	0.006996
attn_gate@11+attn_qkv@11	62.1419	0.171846	[ "blk.11.attn_gate.weight", "blk.11.attn_qkv.weight" ]	2.5224	0.006217
attn_output@11	62.761	0.1776	[ "blk.11.attn_output.weight" ]	3.1415	0.011971
ffn_down@11	59.9345	0.16673	[ "blk.11.ffn_down.weight" ]	0.315	0.001101
ffn_gate@11+ffn_up@11	60.1872	0.165522	[ "blk.11.ffn_gate.weight", "blk.11.ffn_up.weight" ]	0.5677	-0.000107
attn_gate@12+attn_qkv@12	59.1304	0.166325	[ "blk.12.attn_gate.weight", "blk.12.attn_qkv.weight" ]	-0.4891	0.000696
attn_output@12	59.2574	0.16422	[ "blk.12.attn_output.weight" ]	-0.3621	-0.001409
ffn_down@12	59.2901	0.166526	[ "blk.12.ffn_down.weight" ]	-0.3294	0.000897
ffn_gate@12+ffn_up@12	59.1333	0.165666	[ "blk.12.ffn_gate.weight", "blk.12.ffn_up.weight" ]	-0.4862	0.000037
attn_gate@13+attn_qkv@13	60.4704	0.165469	[ "blk.13.attn_gate.weight", "blk.13.attn_qkv.weight" ]	0.8509	-0.00016
attn_output@13	58.8443	0.168696	[ "blk.13.attn_output.weight" ]	-0.7752	0.003067
ffn_down@13	59.5142	0.166005	[ "blk.13.ffn_down.weight" ]	-0.1053	0.000376
ffn_gate@13+ffn_up@13	59.406	0.168074	[ "blk.13.ffn_gate.weight", "blk.13.ffn_up.weight" ]	-0.2135	0.002445
attn_gate@14+attn_qkv@14	60.0529	0.166041	[ "blk.14.attn_gate.weight", "blk.14.attn_qkv.weight" ]	0.4334	0.000412
attn_output@14	62.3683	0.168285	[ "blk.14.attn_output.weight" ]	2.7488	0.002656
ffn_down@14	60.0463	0.16591	[ "blk.14.ffn_down.weight" ]	0.4268	0.000281
ffn_gate@14+ffn_up@14	59.8849	0.166535	[ "blk.14.ffn_gate.weight", "blk.14.ffn_up.weight" ]	0.2654	0.000906
attn_gate@15+attn_qkv@15	59.8054	0.167722	[ "blk.15.attn_gate.weight", "blk.15.attn_qkv.weight" ]	0.1859	0.002093
attn_output@15	59.9846	0.17018	[ "blk.15.attn_output.weight" ]	0.3651	0.004551
ffn_down@15	59.8349	0.166581	[ "blk.15.ffn_down.weight" ]	0.2154	0.000952
ffn_gate@15+ffn_up@15	59.2967	0.166445	[ "blk.15.ffn_gate.weight", "blk.15.ffn_up.weight" ]	-0.3228	0.000816
attn_gate@16+attn_qkv@16	59.0414	0.166336	[ "blk.16.attn_gate.weight", "blk.16.attn_qkv.weight" ]	-0.5781	0.000707
attn_output@16	59.7461	0.167928	[ "blk.16.attn_output.weight" ]	0.1266	0.002299
ffn_down@16	59.3566	0.166838	[ "blk.16.ffn_down.weight" ]	-0.2629	0.001209
ffn_gate@16+ffn_up@16	59.6802	0.167365	[ "blk.16.ffn_gate.weight", "blk.16.ffn_up.weight" ]	0.0607	0.001736
attn_gate@17+attn_qkv@17	59.674	0.166073	[ "blk.17.attn_gate.weight", "blk.17.attn_qkv.weight" ]	0.0545	0.000444
attn_output@17	59.5826	0.166732	[ "blk.17.attn_output.weight" ]	-0.0369	0.001103
ffn_down@17	59.4036	0.166026	[ "blk.17.ffn_down.weight" ]	-0.2159	0.000397
ffn_gate@17+ffn_up@17	59.6345	0.166806	[ "blk.17.ffn_gate.weight", "blk.17.ffn_up.weight" ]	0.015	0.001177
attn_gate@18+attn_qkv@18	59.5952	0.16687	[ "blk.18.attn_gate.weight", "blk.18.attn_qkv.weight" ]	-0.0243	0.001241
attn_output@18	60.1303	0.167422	[ "blk.18.attn_output.weight" ]	0.5108	0.001793
ffn_down@18	59.4116	0.166542	[ "blk.18.ffn_down.weight" ]	-0.2079	0.000913
ffn_gate@18+ffn_up@18	59.9388	0.167293	[ "blk.18.ffn_gate.weight", "blk.18.ffn_up.weight" ]	0.3193	0.001664
attn_gate@19+attn_qkv@19	59.5415	0.169565	[ "blk.19.attn_gate.weight", "blk.19.attn_qkv.weight" ]	-0.078	0.003936
attn_output@19	60.1382	0.167659	[ "blk.19.attn_output.weight" ]	0.5187	0.00203
ffn_down@19	59.6474	0.165869	[ "blk.19.ffn_down.weight" ]	0.0279	0.00024
ffn_gate@19+ffn_up@19	59.5925	0.167246	[ "blk.19.ffn_gate.weight", "blk.19.ffn_up.weight" ]	-0.027	0.001617
attn_gate@2+attn_qkv@2	58.5664	0.167234	[ "blk.2.attn_gate.weight", "blk.2.attn_qkv.weight" ]	-1.0531	0.001605
attn_output@2	60.4142	0.162107	[ "blk.2.attn_output.weight" ]	0.7947	-0.003522
ffn_down@2	59.1256	0.169092	[ "blk.2.ffn_down.weight" ]	-0.4939	0.003463
ffn_gate@2+ffn_up@2	58.8484	0.169829	[ "blk.2.ffn_gate.weight", "blk.2.ffn_up.weight" ]	-0.7711	0.0042
attn_gate@20+attn_qkv@20	59.7211	0.165871	[ "blk.20.attn_gate.weight", "blk.20.attn_qkv.weight" ]	0.1016	0.000242
attn_output@20	59.3605	0.165736	[ "blk.20.attn_output.weight" ]	-0.259	0.000107
ffn_down@20	59.7061	0.16583	[ "blk.20.ffn_down.weight" ]	0.0866	0.000201
ffn_gate@20+ffn_up@20	59.7592	0.166115	[ "blk.20.ffn_gate.weight", "blk.20.ffn_up.weight" ]	0.1397	0.000486
attn_gate@21+attn_qkv@21	59.7188	0.165797	[ "blk.21.attn_gate.weight", "blk.21.attn_qkv.weight" ]	0.0993	0.000168
attn_output@21	59.6201	0.165718	[ "blk.21.attn_output.weight" ]	0.0006	0.000089
ffn_down@21	59.6114	0.165782	[ "blk.21.ffn_down.weight" ]	-0.0081	0.000153
ffn_gate@21+ffn_up@21	59.7143	0.166036	[ "blk.21.ffn_gate.weight", "blk.21.ffn_up.weight" ]	0.0948	0.000407
attn_gate@22+attn_qkv@22	59.6749	0.165808	[ "blk.22.attn_gate.weight", "blk.22.attn_qkv.weight" ]	0.0554	0.000179
attn_output@22	59.6876	0.165776	[ "blk.22.attn_output.weight" ]	0.0681	0.000147
ffn_down@22	59.646	0.165684	[ "blk.22.ffn_down.weight" ]	0.0265	0.000055
ffn_gate@22+ffn_up@22	59.6731	0.166007	[ "blk.22.ffn_gate.weight", "blk.22.ffn_up.weight" ]	0.0536	0.000378
attn_gate@23+attn_qkv@23	59.7053	0.16685	[ "blk.23.attn_gate.weight", "blk.23.attn_qkv.weight" ]	0.0858	0.001221
attn_output@23	59.7115	0.166461	[ "blk.23.attn_output.weight" ]	0.092	0.000832
ffn_down@23	59.6002	0.165815	[ "blk.23.ffn_down.weight" ]	-0.0193	0.000186
ffn_gate@23+ffn_up@23	59.6046	0.165844	[ "blk.23.ffn_gate.weight", "blk.23.ffn_up.weight" ]	-0.0149	0.000215
attn_gate@24+attn_qkv@24	59.6352	0.165663	[ "blk.24.attn_gate.weight", "blk.24.attn_qkv.weight" ]	0.0157	0.000034
attn_output@24	59.7712	0.165952	[ "blk.24.attn_output.weight" ]	0.1517	0.000323
ffn_down@24	59.6361	0.165733	[ "blk.24.ffn_down.weight" ]	0.0166	0.000104
ffn_gate@24+ffn_up@24	59.4662	0.165764	[ "blk.24.ffn_gate.weight", "blk.24.ffn_up.weight" ]	-0.1533	0.000135
attn_gate@25+attn_qkv@25	59.547	0.16562	[ "blk.25.attn_gate.weight", "blk.25.attn_qkv.weight" ]	-0.0725	-0.000009
attn_output@25	59.7157	0.165879	[ "blk.25.attn_output.weight" ]	0.0962	0.00025
ffn_down@25	59.5871	0.165663	[ "blk.25.ffn_down.weight" ]	-0.0324	0.000034
ffn_gate@25+ffn_up@25	59.652	0.165731	[ "blk.25.ffn_gate.weight", "blk.25.ffn_up.weight" ]	0.0325	0.000102
attn_gate@26+attn_qkv@26	59.4355	0.165604	[ "blk.26.attn_gate.weight", "blk.26.attn_qkv.weight" ]	-0.184	-0.000025
attn_output@26	59.701	0.165857	[ "blk.26.attn_output.weight" ]	0.0815	0.000228
ffn_down@26	59.7099	0.165895	[ "blk.26.ffn_down.weight" ]	0.0904	0.000266
ffn_gate@26+ffn_up@26	59.6856	0.166257	[ "blk.26.ffn_gate.weight", "blk.26.ffn_up.weight" ]	0.0661	0.000628
attn_gate@27+attn_qkv@27	59.6696	0.166077	[ "blk.27.attn_gate.weight", "blk.27.attn_qkv.weight" ]	0.0501	0.000448
attn_output@27	59.7747	0.166347	[ "blk.27.attn_output.weight" ]	0.1552	0.000718
ffn_down@27	59.6803	0.16581	[ "blk.27.ffn_down.weight" ]	0.0608	0.000181
ffn_gate@27+ffn_up@27	60.2065	0.166645	[ "blk.27.ffn_gate.weight", "blk.27.ffn_up.weight" ]	0.587	0.001016
attn_gate@3+attn_qkv@3	57.9736	0.170022	[ "blk.3.attn_gate.weight", "blk.3.attn_qkv.weight" ]	-1.6459	0.004393
attn_output@3	60.2641	0.175083	[ "blk.3.attn_output.weight" ]	0.6446	0.009454
ffn_down@3	60.4429	0.167953	[ "blk.3.ffn_down.weight" ]	0.8234	0.002324
ffn_gate@3+ffn_up@3	60.5713	0.167418	[ "blk.3.ffn_gate.weight", "blk.3.ffn_up.weight" ]	0.9518	0.001789
attn_gate@4+attn_qkv@4	60.102	0.165481	[ "blk.4.attn_gate.weight", "blk.4.attn_qkv.weight" ]	0.4825	-0.000148
attn_output@4	59.9959	0.167982	[ "blk.4.attn_output.weight" ]	0.3764	0.002353
ffn_down@4	59.5026	0.164812	[ "blk.4.ffn_down.weight" ]	-0.1169	-0.000817
ffn_gate@4+ffn_up@4	59.1226	0.170794	[ "blk.4.ffn_gate.weight", "blk.4.ffn_up.weight" ]	-0.4969	0.005165
attn_gate@5+attn_qkv@5	60.4099	0.167059	[ "blk.5.attn_gate.weight", "blk.5.attn_qkv.weight" ]	0.7904	0.00143
attn_output@5	59.0201	0.17212	[ "blk.5.attn_output.weight" ]	-0.5994	0.006491
ffn_down@5	60.0891	0.16709	[ "blk.5.ffn_down.weight" ]	0.4696	0.001461
ffn_gate@5+ffn_up@5	60.2806	0.173736	[ "blk.5.ffn_gate.weight", "blk.5.ffn_up.weight" ]	0.6611	0.008107
attn_gate@6+attn_qkv@6	58.0189	0.169133	[ "blk.6.attn_gate.weight", "blk.6.attn_qkv.weight" ]	-1.6006	0.003504
attn_output@6	59.5431	0.168797	[ "blk.6.attn_output.weight" ]	-0.0764	0.003168
ffn_down@6	60.3005	0.169997	[ "blk.6.ffn_down.weight" ]	0.681	0.004368

End of preview.

MERNIK Gnom damage labels

What is Gnom. Gnom is a tiny neural ranker that learns which layers matter. Instead of guessing importance from proxy signals, it trains on these files and then tells the quantizer where to spend bits: important layers stay precise, unimportant ones get compressed harder. You can train your own models the same way — to tell important layers from unimportant ones on any architecture: measure per-layer damage with the sweep script, train the ranker, plug it into your allocator. The bar we hold: transfer to an unseen model (LOMO), not memorization of one.

Where Gnom sits. Three ways to know a tensor's importance: proxy signals (imatrix, Fisher — fast but indirect), full sweep (exact but ~~336 benchmark runs per model), or Gnom: sweep once on a few models, train a tiny ranker (~~150 params, CPU, seconds), rank unseen models with zero sweeps. Without Gnom, MERNIK is craft per model; with it, a pipeline (hours, not weeks).

imatrix calibration → damage sweep (one-time) → Gnom ranker (LOMO-validated)
  → SMSE/Huber allocator (budget-dependent) → preflight + HOLE/TERMINAL
  → loud watchdog → two-ring validation (KLD fast, HE slow)

Why. A static allocator ranks tensors with proxy signals (imatrix, Fisher). Gnom learns the ranking from measured damage instead: take one tensor (or tied group), drop it one tier, measure what breaks. These files are those measurements — the training set behind the Gnom ranker, which cleared the transfer bar (moeshared LOMO = 0.464 vs 0.45, rank LOMO-acc = 0.703 vs 0.60) only after the 1.7B-Q2 tail labels below landed.

How each row was made. group_damage_sweep.py: quantize a copy of the model with exactly one unit pinned one tier lower, then score it. Two damage flavors:

damage — PPL damage on wikitext (cheap, sharpenable, triage only)
damage_kld — KLD vs the unperturbed model on a fixed span (fidelity-sensitive, the column the allocator actually listens to)

What's inside.

File	Rows	Content
`damage_4b_kld_q2.jsonl` / `damage_4b_kld_q3.jsonl` / `damage_4b_kld.jsonl`	20/20/20	Spark-X2.5-4B, paired Q4/Q3/Q2 × 19 units + baseline. The first design with power (within-unit across-tier)
`damage_17_q2_kld.jsonl` / `damage_17_kld.jsonl`	113/113	Spark-1.7B Q2 (+Q3) KLD — the tail labels that moved LOMO over the bar
`damage_17_rlmini.jsonl`	113	Spark-1.7B PPL damage on `rl_reasoning_mini` (343 prompts, 5 RL-oss subsets) + extra KLD columns (`kld_max`, `kld_median`, `dp_rms`, `same_top`)
`damage_4b_rlmini.jsonl`	145	Spark-4B RL-mini PPL damage, all 144 units, zero holes
`damage_m2_rlmini.jsonl`	54	RINIQ-M2 RL-mini (partial leg)
`damage_4b.jsonl` / `damage_minicpm.jsonl` / `damage_v3_q5q2.jsonl` / `damage_kaggle.jsonl`	legacy	early PPL sweeps (kept for provenance, not training)
`*.failures.jsonl`	—	HOLE records (transient, retried) — the failures ledger, not labels
`pool*.json`	—	assembled multi-tier training pools (v2/v3 + RL variants, n up to 333)

Source models (all local GGUF sweeps, imatrix-guided where noted):

damage_4b* — Spark-X2.5-4B (XHToken/Spark-X2.5-4B), 19 units + baseline, Q5_K base
damage_17* — Spark-1.7B, 112 units + baseline, Q5_K base (Q2 leg = the tail)
damage_m2_rlmini — RINIQ-M2 (OxCoder-9B MERNIK build), partial leg
damage_minicpm / damage_v3_q5q2 — MiniCPM5-2B, 168 units
damage_kaggle / gmsd_17_q4 — early probes, kept for provenance

Censored, not missing. TERMINAL units (physically unmeasurable at a tier) carry "damage": null, "retryable": false with a reason code — never imputed, never retried, dropped from training by construction (train_gnom.py guard).

Use. git clone + point gnom/train_gnom.py / train_rank at any pool file. Reproduces the published LOMO numbers on CPU, no GPU needed. Cost of the labels themselves: these files took literally forever — weeks of consumer-GPU sweeps (GTX 1070 + a 16 GB forge box), hundreds of full benchmark runs, two dead servers, one reboot-eaten night. Respect the rows.

What's next. More RL-mini-style labels: reasoning-task damage is worth far more than wiki damage per row, so future sweeps go there. The newest RL-mini KLD sets live in the spotlight repo: wepiqx/MERNIK-rlmini-kld.

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

Downloads last month: 114

Total file size:

199 kB