Chang-chih/leviathan-10.5b-final warn
Weights only ship in a format that can run code when loaded; its tokenizer differs from its claimed base model; glitch tokens that can silently corrupt ordinary input. Plus 1 minor note.
Could not load this model from the Hugging Face API (private, gated, or nonexistent). Findings below are from our archive.
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 2026-08-26102,400-token embedding scanned · 4131 undertrained · 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.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.
low Padding token is the EOS token
The pad token and the (only) EOS token are the same. Fine-tuning frameworks mask pad positions out of the loss, so training on this checkpoint teaches the model to never emit EOS — the Phi-4 / Qwen 2.5 / DeepSeek R1 infinite-generation bug. Safe to serve, hazardous to fine-tune; repoint pad_token at a dedicated token first.
How to fixingot patch
Align the stop-token declarations — a pure metadata fix to `generation_config.json` (and `config.json`).
- Identify the token the chat template actually ends assistant turns with (e.g. `<|eot_id|>`, `<end_of_turn>`, `<|im_end|>`) and make sure its id is in `generation_config.json`'s `eos_token_id` list.
- Keep `config.json`'s `eos_token_id` consistent with (or a subset of) `generation_config.json`'s — runtimes differ in which file they read.
- For the pad-equals-EOS hazard: repoint `pad_token` at a dedicated padding token before fine-tuning; serving is unaffected.
- Until the repo is fixed, pass explicit stop tokens to your serving stack (e.g. vLLM `stop_token_ids`, llama.cpp `--override-kv tokenizer.ggml.eos_token_id`).
medium Vocabulary size differs from claimed parent (102400 vs 32256)
A changed vocab means changed tokenization: strings will split differently than on Chang-chih/leviathan-chat-7b, 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 4131 undertrained tokens (norm < 0.3× the vocabulary median of 8.416), including 1040 plain-ASCII strings that can appear in ordinary input as identifiers — e.g. "IconSuccessEncoded", "IconErrorEncoded", "allClassesLink", "ExternalTaskPojo", "Irefn", "orangehilldev", "navBarCell", "memSeparator". 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.bin (no code executed) found only standard serialization globals (5 distinct, all torch/collections/numpy). Pickle remains an executable format — this verifies the current bytes, not future uploads; prefer a safetensors release.
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 Chang-chih/leviathan-10.5b-final
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 · 40 layers · 4096-dim |
| vocabulary | 102,400 tokens |
| license | apache-2.0 |
| serialization | no safetensors pickle |
| chat template | present (chat_template.jinja) · sha256:8aeba567270fa9a8 |
| claimed lineage | Chang-chih/leviathan-chat-7b, Chang-chih/leviathan-16B-final |
| lineage verified | unverified — weights battery pending |
| glitch-token surface | 4,131 undertrained candidates, 1,040 plain-ASCII |
Full measured fingerprint
| architectures | LlamaForCausalLM |
| library | transformers |
| pipeline | text-generation |
| repo files | 11 — pickle: pytorch_model.bin |
| revision | 0e8dacd85df1 |
| HF snapshot | 108 downloads · 0 likes · updated 2026-07-17 · captured 2026-08-25 |
| pickle audit | pytorch_model.bin — 5 standard global(s) |
| embedding tensor | model.embed_tokens.weight · BF16 · 102,400×4096 |
| embedding norms | median 8.4155 · mean 7.8056 |
| lineage check | parent weights unreadable (repo ships no scannable weights (no safetensors and no pickle checkpoints — GGUF/CoreML/other formats)) |
| glitch-token samples | "IconSuccessEncoded", "IconErrorEncoded", "allClassesLink", "ExternalTaskPojo", "Irefn", "orangehilldev", "navBarCell", "memSeparator", "typeNameLink", "OnSearchSelect", "cachedSer", "Supamiu" |
Battery runs (1)the run trace behind the findings — what each job measured
| battery | status | queued | duration | attempts |
|---|---|---|---|---|
| weights | complete | 2026-08-25 21:53 | 3m | 1 |
weights run 2026-08-25 — measurements
| probes run | glitch-norm-scan, zero-template-token-scan, pickle-static-analysis, lineage-norm-correlation |
| embedding tensor | model.embed_tokens.weight · BF16 · 102,400×4096 |
| glitch surface | 4,131 undertrained, 1,040 plain-ASCII |
| lineage check | not checked (parent weights unreadable (repo ships no scannable weights (no safetensors and no pickle checkpoints — GGUF/CoreML/other formats))) |
Verdict badge
Ship the verdict in your README — it always shows the latest published analysis:
[](https://ingot.tools/models/Chang-chih/leviathan-10.5b-final)