reaperdoesntknow/Qemma-sft warn
Weights only ship in a format that can run code when loaded; Chat template ends turns with <end_of_turn>, which is not a configured stop token; its license differs from its base model's. Plus 2 more issues.
claims base: google/gemma-3-1b-it, Qwen/Qwen3-0.6B · chat template: present · view on Hugging Face ↗
Scan coverageStatic battery2026-08-25Weights battery2026-08-25Behavioral batterynot rundetails
| Battery | Looks at | Status |
|---|---|---|
| Static battery | Metadata & packaging | complete 2026-08-25 |
| Weights battery | Weights forensics: no GPU, no download | complete 2026-08-25262,149-token embedding scanned · 0 undertrained · lineage inconclusive · pickle audit clean |
| Behavioral battery | Live-inference differentials | not run |
Ingot runs three batteries against a model. What each one checks →
Findings
Scanned 2026-08-25 · published from a community scan.
medium Pickle-serialized weights, no safetensors
Weights ship only as pickle-based files (pytorch_model.bin, training_args.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 Chat template ends turns with <end_of_turn>, which is not a configured stop token
The chat template terminates assistant turns with <end_of_turn>, but the effective EOS set (config.json ∪ generation_config.json = [1] → ["<eos>"]) never stops on it. Config-honoring runtimes generate past the terminator until the token budget is exhausted — runaway cost and self-continuing fake turns. Add <end_of_turn>'s id to generation_config.json's eos_token_id.
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 License differs from claimed parent (osl-3.0 vs gemma)
This model declares osl-3.0 while its claimed base google/gemma-3-1b-it declares gemma. 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 Stop token dropped vs claimed parent but still used by the template
google/gemma-3-1b-it stops on [1,106]; this repo only stops on [1], losing <end_of_turn> (106) — which this repo's own chat template still emits. Config-honoring runtimes will generate straight past it. Restore the lost id(s) to generation_config.json's eos_token_id.
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 (262149 vs 262144)
A changed vocab means changed tokenization: strings will split differently than on google/gemma-3-1b-it, 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.
info Embedding-norm glitch scan clean
No undertrained tokens found: every non-special token's embedding norm is above 0.3× the vocabulary median (1.244). The glitch-token data-corruption class has no candidate surface in this model.
info Pickle static analysis clean
Opcode-level parse of pytorch_model.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.
low Lineage vs claimed parent google/gemma-3-1b-it inconclusive
Mean embedding-row cosine similarity to the declared base is 0.758 — below the 0.8 typical of true derivatives but not low enough to call mislabeled. Heavy continued pretraining or vocabulary surgery can look like this; verify provenance before relying on the parent's safety or licensing posture.
How to fix
Fix or verify the `base_model` declaration so lineage checks can run.
- If you own the repo: correct the `base_model` field in the model card metadata to the real, public parent.
- If you don't: identify the true parent (config architecture + weight shapes narrow it fast) and re-scan with that lineage in mind.
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 reaperdoesntknow/Qemma-sft
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 2026-08-25.
| architecture | gemma3_text · 26 layers · 1152-dim |
| vocabulary | 262,149 tokens |
| license | osl-3.0 |
| serialization | no safetensors pickle |
| chat template | present (chat_template.jinja) · sha256:7de1c58e208eda46 |
| claimed lineage | google/gemma-3-1b-it, Qwen/Qwen3-0.6B |
| lineage verified | inconclusive vs google/gemma-3-1b-it — embedding-row cosine 0.758 |
| glitch-token surface | clean no undertrained tokens |
Full measured fingerprint
| architectures | Gemma3ForCausalLM |
| library | transformers |
| pipeline | text-generation |
| repo files | 15 — pickle: pytorch_model.bin, training_args.bin |
| revision | f873068b4b54 |
| HF snapshot | 3.0k downloads · 0 likes · updated 2026-08-18 · captured 2026-08-25 |
| pickle audit | pytorch_model.bin — 3 standard global(s) |
| embedding tensor | model.embed_tokens.weight · F32 · 262,149×1152 |
| embedding norms | median 1.2435 · mean 1.2397 |
| lineage check | inconclusive — cosine 0.7585 over 64 sampled rows vs google/gemma-3-1b-it |
Battery runs (1)the run trace behind the findings — what each job measured
| battery | status | queued | duration | attempts |
|---|---|---|---|---|
| weights | complete | 2026-08-25 21:46 | 81s | 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 · F32 · 262,149×1152 |
| glitch surface | 0 undertrained, 0 plain-ASCII |
| lineage check | inconclusive — cosine 0.7585 over 64 rows vs google/gemma-3-1b-it |
Verdict badge
Ship the verdict in your README — it always shows the latest published analysis:
[](https://ingot.tools/models/reaperdoesntknow/Qemma-sft)