Model page

Qwen/Qwen3.5-9B warn

downloads 13.7Mlikes 1.8klicense apache-2.0arch qwen3_5params 9653.1Mupdated 2026-03-02

claims base: Qwen/Qwen3.5-9B-Base · 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 248320 tokens (BF16, 4096-dim) found 1585 undertrained candidates, 397 plain-ASCII. Weights battery: embedding-norm scan over 248320 tokens (BF16, 4096-dim) found 1585 undertrained candidates, 397 plain-ASCII. Lineage vs Qwen/Qwen3.5-9B-Base: consistent. Scanned 2026-08-20 (published from a community scan).

medium Undertrained (glitch) token surface in vocabulary

Embedding-norm scan flagged 1585 undertrained tokens (norm < 0.3× the vocabulary median of 0.886), including 397 plain-ASCII strings that can appear in ordinary input as identifiers — e.g. "tedothi", "ForCanBeConvertedToF", "ForCanBeConverted", "szexf", "Kinhted", "xfabl", "PostalCodesNL", "useRalative". 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 Qwen/Qwen3.5-9B-Base

Mean cosine similarity of 64 sampled token-embedding rows against Qwen/Qwen3.5-9B-Base is 0.999 — 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.

architectureqwen3_5
parameters9653.1M
vocabulary248,320 tokens
licenseapache-2.0
serializationsafetensors
chat templatepresent · sha256:a4aee8afcf2e0711
claimed lineageQwen/Qwen3.5-9B-Base
lineage verifiedconsistent vs Qwen/Qwen3.5-9B-Base — embedding-row cosine 0.999
glitch-token surface1,585 undertrained candidates, 397 plain-ASCII
Full fingerprint
architecturesQwen3_5ForConditionalGeneration
librarytransformers
pipelineimage-text-to-text
repo files16
revisionc20223623576
HF snapshot13.6M downloads · 1.8k likes · updated 2026-03-02 · captured 2026-08-20
embedding tensormodel.language_model.embed_tokens.weight · BF16 · 248,320×4096
embedding normsmedian 0.8862 · mean 0.8524
lineage checkconsistent — cosine 0.9993 over 64 sampled rows vs Qwen/Qwen3.5-9B-Base
glitch-token samples"tedothi", "ForCanBeConvertedToF", "ForCanBeConverted", "szexf", "Kinhted", "xfabl", "PostalCodesNL", "useRalative", "tarsker", "tarskereso", "ejahter", "useRal"

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/Qwen/Qwen3.5-9B/badge.svg)](https://ingot.tools/models/Qwen/Qwen3.5-9B)
Gate it in CI