Model page

dphn/dolphin-2.9.1-yi-1.5-34b warn

downloads 4.6Mlikes 65license apache-2.0arch llamaparams 34388.9Mupdated 2025-09-08

claims base: 01-ai/Yi-1.5-34B · chat template: present · view on Hugging Face ↗

Ingot findings

Static battery clean: safetensors weights, license declared, no template/tokenizer drift detected. Deep battery not yet run. Weights battery: embedding-norm scan over 64000 tokens (BF16, 7168-dim) found 1149 undertrained candidates, 756 plain-ASCII. Lineage vs 01-ai/Yi-1.5-34B: consistent. Scanned 2026-08-20 (published from a community scan).

medium Undertrained (glitch) token surface in vocabulary

Embedding-norm scan flagged 1149 undertrained tokens (norm < 0.3× the vocabulary median of 0.916), including 756 plain-ASCII strings that can appear in ordinary input as identifiers — e.g. "<|unused091|>", "<|unused121|>", "<|unused123|>", "<|unused114|>", "<|unused022|>", "<|unused030|>", "<|unused100|>", "<|unused102|>". 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 GPU deep 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.

  1. 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.
  2. 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.
  3. 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 01-ai/Yi-1.5-34B

Mean cosine similarity of 64 sampled token-embedding rows against 01-ai/Yi-1.5-34B is 1.000 — the weights plausibly descend from the declared base (relation: unspecified).

How to fix

Fix or verify the `base_model` declaration so lineage checks can run.

  1. If you own the repo: correct the `base_model` field in the model card metadata to the real, public parent.
  2. If you don't: identify the true parent (config architecture + weight shapes narrow it fast) and re-scan with that lineage in mind.

Remediation guidance addresses the documented findings only. It is evidence-driven repair, not a safety certification of the model.

Fingerprint

The durable weights-and-metadata profile of this model, rebuilt on every scan and deep-battery run. Updated 2026-08-20.

architecturellama · 60 layers · 7168-dim
parameters34388.9M
vocabulary64,000 tokens
licenseapache-2.0
serializationsafetensors
chat templatepresent · sha256:58c1a1f04baa7ada
claimed lineage01-ai/Yi-1.5-34B
lineage verifiedconsistent vs 01-ai/Yi-1.5-34B — embedding-row cosine 1.000
glitch-token surface1,149 undertrained candidates, 756 plain-ASCII
Full fingerprint
architecturesLlamaForCausalLM
librarytransformers
pipelinetext-generation
repo files24
revision0141cba238d0
HF snapshot4.6M downloads · 65 likes · updated 2025-09-08 · captured 2026-08-20
embedding tensormodel.embed_tokens.weight · BF16 · 64,000×7168
embedding normsmedian 0.9163 · mean 0.881
lineage checkconsistent — cosine 0.9996 over 64 sampled rows vs 01-ai/Yi-1.5-34B
glitch-token samples"<|unused091|>", "<|unused121|>", "<|unused123|>", "<|unused114|>", "<|unused022|>", "<|unused030|>", "<|unused100|>", "<|unused102|>", "<|unused152|>", "<|unused140|>", "<|unused099|>", "<|unused002|>"

Verdict badge

Ship the verdict in your README — it always shows the latest published analysis:

Ingot verdict: warn

[![Ingot scan](https://ingot.tools/api/v1/models/dphn/dolphin-2.9.1-yi-1.5-34b/badge.svg)](https://ingot.tools/models/dphn/dolphin-2.9.1-yi-1.5-34b)
Gate it in CI