0arch-io/dolphin-v2-8b-abliterated warn
The chat template was dropped from its base model, which changes behavior; glitch tokens that can silently corrupt ordinary input. Plus 1 minor note.
claims base: Qwen/Qwen3-8B · chat template: not found · 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 2026-08-21151,936-token embedding scanned · 3000 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 Chat template dropped vs parent
Qwen/Qwen3-8B ships a chat template; this repo does not. Serving stacks will silently fall back to a generic template, changing behavior. (In our 296-model census, 78% of pure quantization re-releases changed or dropped the template.)
How to fixingot patch
Restore the parent's chat template in `tokenizer_config.json` — a pure metadata fix.
- Run `ingot patch <owner/model>` — the patch manifest carries the parent's template and applies it to a local copy's `tokenizer_config.json`.
- Or fix by hand: copy the `chat_template` value from the parent repo's `tokenizer_config.json` into this model's, and pin your serving stack to that file.
- If the drift was intentional (the author retrained on a new template), confirm that in the model card before "fixing" it — restoring the parent template on retrained weights changes behavior too.
medium Undertrained (glitch) token surface in vocabulary
Embedding-norm scan flagged 3000 undertrained tokens (norm < 0.3× the vocabulary median of 1.407), including 102 plain-ASCII strings that can appear in ordinary input as identifiers — e.g. "$PostalCodesNL", "ForCanBeConvertedToF", "PostalCodesNL", "ForCanBeConverted", "useRalative", "thuisontvangst", "useRal", "sexkontakte". 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 Qwen/Qwen3-8B
Mean cosine similarity of 64 sampled token-embedding rows against Qwen/Qwen3-8B is 0.998 — the weights plausibly descend from the declared base (relation: unspecified).
low Undertrained tokens present; echo probe skipped (no chat template)
3000 undertrained tokens found (16 ASCII candidates) but the repo ships no chat template, so the behavioral echo probe was skipped. Treat the candidates as unverified risk surface.
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.
Fix it
Some findings are metadata-level and patchable — apply the fixes to your local copy (your weights never leave your machine):
npx @ingotai/scan patch 0arch-io/dolphin-v2-8b-abliterated
Remediation guidance addresses the documented findings only. It is evidence-driven repair, not a safety certification of the model.
Fingerprint
The durable profile of this model: measured weights-and-metadata facts, rebuilt on every scan and battery run. Updated 2026-08-21.
| architecture | qwen3 · 36 layers · 4096-dim |
| parameters | 8190.7M |
| vocabulary | 151,936 tokens |
| license | apache-2.0 |
| serialization | safetensors + gguf pickle |
| chat template | none |
| claimed lineage | Qwen/Qwen3-8B |
| lineage verified | consistent vs Qwen/Qwen3-8B — embedding-row cosine 0.998 |
| glitch-token surface | 3,000 undertrained candidates, 102 plain-ASCII |
Full measured fingerprint
| architectures | Qwen3ForCausalLM |
| pipeline | text-generation |
| repo files | 16 — pickle: refusal_direction.pt, refusal_direction_0_layer35.pt, refusal_direction_1_layer34.pt, refusal_direction_2_layer36.pt, refusal_direction_3_layer33.pt, refusal_direction_4_layer16.pt |
| revision | 527a037d0815 |
| HF snapshot | 857 downloads · 3 likes · updated 2026-02-24 · captured 2026-08-21 |
| embedding tensor | model.embed_tokens.weight · BF16 · 151,936×4096 |
| embedding norms | median 1.4069 · mean 1.3338 |
| lineage check | consistent — cosine 0.9984 over 64 sampled rows vs Qwen/Qwen3-8B |
| glitch-token samples | "$PostalCodesNL", "ForCanBeConvertedToF", "PostalCodesNL", "ForCanBeConverted", "useRalative", "thuisontvangst", "useRal", "sexkontakte", "NdrFc", "webElementX", "sextreffen", "wannonce" |
Battery runs (2)the run trace behind the findings — what each job measured
| battery | status | queued | duration | attempts |
|---|---|---|---|---|
| gpu | complete | 2026-08-27 04:31 | 3m | 1 |
| weights | complete | 2026-08-21 05:13 | 2m | 1 |
gpu run 2026-08-27 — measurements
| probes run | glitch |
| probes skipped | glitch-echo: no chat template |
weights run 2026-08-21 — measurements
| probes run | glitch-norm-scan, lineage-norm-correlation |
| embedding tensor | model.embed_tokens.weight · BF16 · 151,936×4096 |
| glitch surface | 3,000 undertrained, 102 plain-ASCII |
| lineage check | consistent — cosine 0.9984 over 64 rows vs Qwen/Qwen3-8B |
Verdict badge
Ship the verdict in your README — it always shows the latest published analysis:
[](https://ingot.tools/models/0arch-io/dolphin-v2-8b-abliterated)