EldritchLabs/KrakenSakura-Maelstrom-12B-v1 fail
Glitch tokens silently corrupt pipeline records (Chinese); its tokenizer differs from its claimed base model; glitch tokens that can silently corrupt ordinary input. Plus 1 more issue.
claims base: allura-org/Tlacuilo-12B, ChaoticNeutrals/Captain_Eris_Noctis-12B-v0.420, DavidAU/Mistral-Nemo-2407-12B-Thinking-Claude-Gemini-GPT5.2-Uncensored-HERETIC, DavidAU/MN-Dark-Planet-TITAN-12B, DavidAU/MN-GRAND-Gutenberg-Lyra4-Lyra-12B-DARKNESS, DreadPoor/Famino-12B-Model_Stock, EldritchLabs/Cactus-Dream-Horror-12B, EldritchLabs/Human-Like-Mistral-Nemo-Instruct-2407-MPOA, EldritchLabs/Kraken-Karcher-12B-v1, EldritchLabs/MN-12B-Mag-Mell-R1-Uncensored-Scale1.2, EldritchLabs/MN-12B-RP-Ink-Longform-MPOA, Epiculous/Violet_Twilight-v0.2, IggyLux/MN-VelvetCafe-RP-12B-V2, inflatebot/MN-12B-Mag-Mell-R1, LatitudeGames/Muse-12B, LatitudeGames/Wayfarer-2-12B, mistralai/Mistral-Nemo-Instruct-2407, MrRikyz/StarlightMoon-Foxfire-12B, MuXodious/Rocinante-X-12B-v1-absolute-heresy, Naphula/Ancient-Awakening-12B, Naphula/Riemannian-Redshift-12B-v1, ohyeah1/Violet-Lyra-Gutenberg-v2, PocketDoc/Dans-SakuraKaze-V1.0.0-12b, PygmalionAI/Pygmalion-3-12B, rAIfle/Questionable-MN-bf16, ReadyArt/Dark-Nexus-12B-v2.0, ReadyArt/Forgotten-Safeword-12B-v4.0, redrix/GodSlayer-12B-ABYSS, Retreatcost/Chrysologus-12B, Retreatcost/Impish-LongPen-12B, Retreatcost/KansenSakura-Conflagration-RP-12b, SicariusSicariiStuff/Impish_Bloodmoon_12B, Sorihon/Celestial-Queen-12B-Heretic, SuperbEmphasis/MN-12b-RP-Ink-RP-Longform, SuperbEmphasis/Omega-Darker_The-Final-Directive-Longform-Stage2-ERP-12B-v0.2, TheDrummer/Rocinante-X-12B-v1, Vortex5/Aurora-Mirage-12B, Vortex5/Prototype-X-12b · chat template: present · view on Hugging Face ↗
Scan coverageStatic battery2026-08-27Weights battery2026-08-21Behavioral batterycompletedetails
| Battery | Looks at | Status |
|---|---|---|
| Static battery | Metadata & packaging | complete 2026-08-27 |
| Weights battery | Weights forensics: no GPU, no download | complete 2026-08-21131,075-token embedding scanned · 141 undertrained · lineage consistent |
| Behavioral battery | Live-inference differentials | completefull differential battery (curated) |
Ingot runs three batteries against a model. What each one checks →
Findings
Scanned 2026-08-27 · published from a community scan.
medium Vocabulary size differs from claimed parent (131075 vs 131074)
A changed vocab means changed tokenization: strings will split differently than on allura-org/Tlacuilo-12B, which can shift behavior on identifiers, codes, and non-English text.
How to fixweight-level
Not patchable: the vocab size mirrors the embedding matrix in the weights. Verify the change was intentional.
- Do not edit `vocab_size` in config.json to "match the parent" — it must equal the embedding table in the shipped weights or the model won't load.
- Diff the tokenizers (`tokenizer.json` / added_tokens) against the parent to see what was added or removed, and test your own identifiers, codes, and non-English text through both.
- If the drift is unexplained by the model card, treat tokenization-sensitive behavior as unvalidated on this model.
medium Undertrained (glitch) token surface in vocabulary
Embedding-norm scan flagged 141 undertrained tokens (norm < 0.3× the vocabulary median of 0.626), including 24 plain-ASCII strings that can appear in ordinary input as identifiers — e.g. "erresident", "abezian", "komert", "higiez", "pemerint", "banako", "komertzio", "tanleria". In models where this class was tested behaviorally, such tokens silently rewrote user input into confident, schema-valid, wrong output. These are candidates from the weights alone; behavioral confirmation requires the behavioral battery.
How to fixruntime guardweight-level
Keep the affected token strings out of the model's input — the scan-derived runtime guard carries this model's exact blocklist.
- Fetch this model's guard artifact (`/api/v1/guard/<owner>/<model>`): the confirmed corrupting tokens and the low-norm candidate list, derived from the published scan.
- Screen inbound text with it (the `@ingotai/guard` package is a reference implementation) and route flagged records to a different model or human review — verbatim-copy tasks on flagged strings are the failure mode.
- The underlying cause is undertrained embeddings in the weights; a true fix is weight-level (continued pretraining on the affected tokens) — that is not a patch, it's a training job.
info Weights consistent with claimed parent allura-org/Tlacuilo-12B
Mean cosine similarity of 64 sampled token-embedding rows against allura-org/Tlacuilo-12B is 1.000 — the weights plausibly descend from the declared base (relation: unspecified).
low Partial glitch-token echo degradation
Echo failures on 6/16 undertrained tokens vs 0/8 controls — a differential exists but below the confirmation bar (≥50% glitch failures with clean controls).
high Glitch tokens silently corrupt pipeline records (Chinese)
Placed in Chinese-language slots (username, order reference, SKU, verbatim archive), 2 of this model's undertrained Chinese tokens silently corrupted 7 of 8 realistic pipeline runs across 4 scenario types while matched control tokens passed (1 control corruptions), and 2 corruptions stayed schema-valid JSON — the record is wrong but nothing errors (refusals, where the model declined rather than fabricated, are excluded). E.g. "页面存档" → verbatim_archive_zh produced "编号： ```yaml 编号：1 ```"; "页面存档" → support_confirm_username_zh produced "欢迎注册，your username is "yckyourselfyourselfyourselfyourselfyo"; "页面存档" → crm_sku_json_zh produced "```json { "product_sku": null } ```". Greedy decoding, temperature 0, seed 0.
medium Glitch tokens confirmed behaviorally (echo test, Chinese)
Asked to repeat its own undertrained Chinese tokens verbatim, the model failed on 2/3 while repeating 8/8 matched normal Chinese tokens correctly — e.g. "页面存档" → """ (Note: The string "```" is a common w"; "衆国" → ""ーヨ" (Note: The string "ーヨ" is not a val". Greedy decoding, temperature 0, seed 0.
Remediation guidance addresses the documented findings only. It is evidence-driven repair, not a safety certification of the model.
Check every checkpoint before it ships
Use the web app, API, CLI, or CI gate to scan candidate checkpoints and catch model drift before deployment. Public-model scans publish to the open database; paid plans add the volume needed for continuous checks.
Fingerprint
The durable profile of this model: measured weights-and-metadata facts, rebuilt on every scan and battery run. Updated 2026-08-21.
Full measured fingerprint
| architectures | MistralForCausalLM |
| library | transformers |
| pipeline | text-generation |
| repo files | 19 |
| revision | cfc1f3b8b119 |
| HF snapshot | 305 downloads · 17 likes · updated 2026-04-04 · captured 2026-08-21 |
| embedding tensor | model.embed_tokens.weight · BF16 · 131,075×5120 |
| embedding norms | median 0.6261 · mean 0.6028 |
| lineage check | consistent — cosine 0.9996 over 64 sampled rows vs allura-org/Tlacuilo-12B |
| glitch-token samples | "erresident", "abezian", "komert", "higiez", "pemerint", "banako", "komertzio", "tanleria", "igelts", "miejs", "-usti", "zimendu" |
Battery runs (2)the run trace behind the findings — what each job measured
| battery | status | queued | duration | attempts |
|---|---|---|---|---|
| gpu | complete | 2026-08-27 08:21 | 3m | 1 |
| weights | complete | 2026-08-21 05:15 | 2m | 1 |
gpu run 2026-08-27 — measurements
| probes run | glitch |
weights run 2026-08-21 — measurements
| probes run | glitch-norm-scan, lineage-norm-correlation |
| embedding tensor | model.embed_tokens.weight · BF16 · 131,075×5120 |
| glitch surface | 141 undertrained, 24 plain-ASCII |
| lineage check | consistent — cosine 0.9996 over 64 rows vs allura-org/Tlacuilo-12B |
Verdict badge
Ship the verdict in your README — it always shows the latest published analysis:
[](https://ingot.tools/models/EldritchLabs/KrakenSakura-Maelstrom-12B-v1)