safe049/ParuMaid-Llama3-Chinese-8B warn
Weights only ship in a format that can run code when loaded; its license differs from its base model's; glitch tokens that can silently corrupt ordinary input. Plus 1 more issue.
claims base: safe049/Ruozhiba_llama3 · chat template: present · view on Hugging Face ↗
Scan coverageStatic battery2026-08-26Weights battery2026-08-26Behavioral batterynot rundetails
| Battery | Looks at | Status |
|---|---|---|
| Static battery | Metadata & packaging | complete 2026-08-26 |
| Weights battery | Weights forensics: no GPU, no download | complete 2026-08-26128,256-token embedding scanned · 461 undertrained · lineage consistent · pickle audit clean |
| Behavioral battery | Live-inference differentials | not run |
Ingot runs three batteries against a model. What each one checks →
Findings
Scanned 2026-08-26 · published from a community scan.
medium Pickle-serialized weights, no safetensors
Weights ship only as pickle-based files (pytorch_model-00001-of-00004.bin, pytorch_model-00002-of-00004.bin, pytorch_model-00003-of-00004.bin, …). Loading pickle executes arbitrary code from the file — prefer a safetensors release or load in a sandbox.
How to fix
Convert the weights to safetensors before loading them anywhere that matters.
- Do not load the pickle files in-process — pickle deserialization executes arbitrary code from the file.
- Convert locally in a sandbox: `pip install safetensors` and use `safetensors.torch.save_file` on a state dict loaded with `torch.load(..., weights_only=True)` (refuses most code-execution payloads), or use Hugging Face's `convert.py` space/script.
- Pin the exact revision hash you converted from, and load only your converted safetensors artifact from then on.
medium License differs from claimed parent (apache-2.0 vs llama3)
This model declares apache-2.0 while its claimed base safe049/Ruozhiba_llama3 declares llama3. Verify the re-license is permitted before commercial use.
How to fix
Verify the re-license is actually permitted before relying on it.
- Read the parent's license for derivative-work and re-licensing terms — many open-weight licenses (e.g. Llama-family) do not permit arbitrary re-licensing.
- If the re-license is not permitted, the parent's terms govern your use regardless of what this repo declares.
medium Undertrained (glitch) token surface in vocabulary
Embedding-norm scan flagged 461 undertrained tokens (norm < 0.3× the vocabulary median of 0.601), including 131 plain-ASCII strings that can appear in ordinary input as identifiers — e.g. "TokenNameIdentifier", "ForCanBeConverted", "ForCanBeConvertedToF", "PostalCodesNL", "$PostalCodesNL", "useRalative", "ilmektedir", "CLIIIK". 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 Pickle static analysis clean
Opcode-level parse of pytorch_model-00001-of-00004.bin, pytorch_model-00002-of-00004.bin, pytorch_model-00003-of-00004.bin, pytorch_model-00004-of-00004.bin (no code executed) found only standard serialization globals (3 distinct, all torch/collections/numpy). Pickle remains an executable format — this verifies the current bytes, not future uploads; prefer a safetensors release.
info Weights consistent with claimed parent safe049/Ruozhiba_llama3
Mean cosine similarity of 64 sampled token-embedding rows against safe049/Ruozhiba_llama3 is 1.000 — the weights plausibly descend from the declared base (relation: unspecified).
medium GGUF template ends turns with <|eot_id|> but the file stops on <|end_of_text|>
The embedded chat template terminates assistant turns with <|eot_id|>, while tokenizer.ggml.eos_token_id points at <|end_of_text|> (128001). llama.cpp-family runtimes stop on the metadata EOS, so generation runs past the terminator until the token budget is exhausted. Fix the eos_token_id in the GGUF metadata (gguf-py, no requant needed).
How to fixingot patch
Fix the GGUF's embedded metadata in place with gguf-py — template, EOS id, and pre-tokenizer are all metadata-editable; no requant needed.
- Template or EOS drift: copy the current values from the source repo and write them into the GGUF (`gguf_set_metadata.py` / gguf-py) — the tensor data is untouched.
- Missing pre-tokenizer type: reconvert with a current `convert_hf_to_gguf.py`, or set the correct `tokenizer.ggml.pre` for the architecture.
- Until the file is fixed, override at load time: llama.cpp `--override-kv tokenizer.ggml.eos_token_id=int:<id>` and `--chat-template-file <fixed.jinja>`.
- Prefer a re-upload from the quantizer once the source repo's fix lands — already-downloaded GGUFs never pick up upstream fixes on their own.
low GGUF chat template differs from the source repo's
The template embedded at conversion time (1112 chars) no longer matches the source repo's current template (1142 chars). Source-repo template fixes never propagate into converted GGUFs — 43% of popular GGUF repos drift this way, including quants that resurrect already-fixed launch bugs. Diff the two before deploying; re-embed with gguf-py if the source's fix matters.
How to fixingot patch
Fix the GGUF's embedded metadata in place with gguf-py — template, EOS id, and pre-tokenizer are all metadata-editable; no requant needed.
- Template or EOS drift: copy the current values from the source repo and write them into the GGUF (`gguf_set_metadata.py` / gguf-py) — the tensor data is untouched.
- Missing pre-tokenizer type: reconvert with a current `convert_hf_to_gguf.py`, or set the correct `tokenizer.ggml.pre` for the architecture.
- Until the file is fixed, override at load time: llama.cpp `--override-kv tokenizer.ggml.eos_token_id=int:<id>` and `--chat-template-file <fixed.jinja>`.
- Prefer a re-upload from the quantizer once the source repo's fix lands — already-downloaded GGUFs never pick up upstream fixes on their own.
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 safe049/ParuMaid-Llama3-Chinese-8B
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 2026-08-26.
| architecture | llama |
| vocabulary | 128,256 tokens |
| license | apache-2.0 |
| serialization | gguf pickle |
| chat template | present · sha256:57f3ad9f7395b2e9 |
| claimed lineage | safe049/Ruozhiba_llama3 |
| lineage verified | consistent vs safe049/Ruozhiba_llama3 — embedding-row cosine 1.000 |
| glitch-token surface | 461 undertrained candidates, 131 plain-ASCII |
Full measured fingerprint
| architectures | LlamaForCausalLM |
| library | transformers |
| pipeline | text-generation |
| repo files | 15 — pickle: pytorch_model-00001-of-00004.bin, pytorch_model-00002-of-00004.bin, pytorch_model-00003-of-00004.bin, pytorch_model-00004-of-00004.bin |
| revision | 1c557afc8f49 |
| HF snapshot | 64 downloads · 1 likes · updated 2024-10-27 · captured 2026-08-25 |
| pickle audit | pytorch_model-00001-of-00004.bin, pytorch_model-00002-of-00004.bin, pytorch_model-00003-of-00004.bin, pytorch_model-00004-of-00004.bin — 3 standard global(s) |
| embedding tensor | model.embed_tokens.weight · F16 · 128,256×4096 |
| embedding norms | median 0.6012 · mean 0.5911 |
| lineage check | consistent — cosine 1 over 64 sampled rows vs safe049/Ruozhiba_llama3 |
| glitch-token samples | "TokenNameIdentifier", "ForCanBeConverted", "ForCanBeConvertedToF", "PostalCodesNL", "$PostalCodesNL", "useRalative", "ilmektedir", "CLIIIK", "_ComCallableWrapper", "krvldkf", "webElementXpaths", "useRalativeImagePath" |
Battery runs (1)the run trace behind the findings — what each job measured
| battery | status | queued | duration | attempts |
|---|---|---|---|---|
| weights | complete | 2026-08-25 21:57 | 3m | 1 |
weights run 2026-08-25 — measurements
| probes run | glitch-norm-scan, zero-template-token-scan, pickle-static-analysis, lineage-norm-correlation, gguf-metadata |
| embedding tensor | model.embed_tokens.weight · F16 · 128,256×4096 |
| glitch surface | 461 undertrained, 131 plain-ASCII |
| lineage check | consistent — cosine 1 over 64 rows vs safe049/Ruozhiba_llama3 |
Verdict badge
Ship the verdict in your README — it always shows the latest published analysis:
[](https://ingot.tools/models/safe049/ParuMaid-Llama3-Chinese-8B)