From Source Revision to GGUF: Validating an Unsloth Export with OMIV
The export succeeded, and both current artifact identities matched retained historical C1 size/SHA-256 observations.
C1 was an earlier historical observation for which the artifact sizes and SHA-256 values were retained. The historical bytes themselves were unavailable. OMIV therefore classifies the result as HISTORICAL_C1_MATCH: a match between current identities and retained observations, not a new direct C1/C2 file comparison.
Direct answer to the original question
The tested Unsloth export path successfully produced both the Q8_0 GGUF and BF16 mmproj. In the inspected target-native metadata, GGUF structure and tensor information were observable, but the exact source revision was not.
OMIV did not observe a cryptographic or format-native binding between the main GGUF and mmproj; their association remained contextual. It also recorded tokenizer-configuration changes in the work copy but did not determine which component originated them.
Both current artifact sizes and SHA-256 values matched retained historical C1 observations. Semantic fidelity, numerical quantization fidelity, and runtime compatibility were not evaluated.
Why this case
This case began with a public technical exchange with Daniel Han. We are grateful to Daniel for suggesting Gemma 4 E2B IT as the first reproducible case and for giving OMIV the opportunity to test a real Unsloth export path. OMIV then independently executed and analyzed the export.
That acknowledgement is limited to the public exchange. It does not imply partnership, endorsement, certification, joint authorship, Unsloth approval, or prior review of these results.
Detailed artifact results
The retained source identity was google/gemma-4-E2B-it at revision 3e22461f65e89153144f8adb70e3b8c2cc9845a7. Its nine-file, 10,278,849,571-byte manifest remained unchanged during export.
| Result | Main Q8_0 GGUF | BF16 mmproj |
|---|---|---|
| Format | GGUF v3 | GGUF v3 |
| Bytes | 4,954,594,880 | 986,833,792 |
| SHA-256 | fbf3f93d603e92dbcf00742163cc3018b9bf8fbfbcf87158cc836e5ed0bc9455 |
e00089a95d2dc85c71f249bd83760e348f721757318802ddb889c65ec64eae3c |
| Metadata entries | 51 | 39 |
| Tensors | 601 | 1,411 |
| Tensor types | 318 Q8_0; 283 F32 | 247 BF16; 1,164 F32 |
| Tensor ranges | In bounds; non-overlapping | In bounds; non-overlapping |
Paired deterministic OMIV-native reports were byte-identical. Together, these checks establish the current artifacts' identities, their observed GGUF structure, and the repeatability of OMIV's reports.
How the result was framed
OMIV kept three kinds of statement separate. Current file size, SHA-256, GGUF structure, and tensor-range checks are directly verified facts. Export success and the C1 values are retained historical observations. The relationship between the main GGUF and mmproj is a contextual inference with an explicit strength and boundary.
During the separate offline artifact-analysis phase, OMIV recomputed custody identities, parsed headers, metadata, and tensor descriptors, checked declared counts and ranges, compared deterministic reports, and assembled the available source/work-copy/target lineage. That phase did not load a model, run inference, invoke an exporter, use a GPU, or interpret tensor payload values for fidelity testing.
This separation prevents one kind of evidence from silently standing in for another. A matching size and SHA-256 establishes identity against the retained observation; it does not establish model behavior. A structurally valid GGUF establishes that the inspected structure is internally consistent; it does not establish output quality or runtime correctness.
Lineage and artifact association
Both artifacts exposed useful target-native structure, but neither target-native GGUF metadata inventory contained the exact source revision. The revision remains available in retained evidence, not in the target files themselves. This is a provenance-observability gap, not proof of incorrect weights.
OMIV also recorded tokenizer-configuration work-copy behavior described in the public claims and evidence registries. It did not determine which component originated that behavior, and the observation is not automatically evidence of a defect.
The main GGUF and mmproj have a moderate contextual association through their shared export context and compatible metadata. No cryptographic or format-native binding between them was observed. Their association is therefore contextual rather than cryptographic or format-native; this is an association gap, not proof of incompatibility. The study did not determine which component originated the provenance, association, or tokenizer-lineage behavior.
Why this matters
Producing a file is not the whole model-delivery question. Release teams also need observable source identity, tokenizer and configuration lineage, artifact association, structural validity, output identity, and explicit boundaries around what was not evaluated. Those properties matter when an exported model moves between a publisher, a conversion pipeline, a runtime, hardware, and a deployment review.
OMIV turns those questions into machine-readable claims, evidence references, and limitations. This keeps directly verified facts, retained historical observations, contextual inferences, and unevaluated areas distinct instead of reducing the result to one unsupported pass/fail label.
Who this is relevant to
- Model publishers and release teams
- Conversion and quantization vendors
- Inference-runtime maintainers
- Accelerator and chip vendors
- Enterprises qualifying models for internal deployment
Boundaries
This study did not evaluate semantic fidelity, numerical quantization fidelity, or runtime compatibility. Q8_0 is a representation fact, not a quality result. Structural validity and export success do not show that a particular runtime will load or correctly execute the artifacts.
Published commitments bind the identity of retained private evidence records, but public commitments do not make undisclosed private records independently reviewable. The public materials are not a standalone exporter reproduction recipe.
OMIV remains pre-1.0. It is not a certification authority, safety evaluation, authenticity oracle, or proof of runtime correctness.
Evidence and next cases
The OMIV evidence Dataset provides compact public summaries, claims, evidence references, limitations, and checksums. The GitHub case study provides the canonical technical mapping.
Maintainers with a real model export, conversion, deployment, or runtime path are invited to propose another evidence-backed case through the OMIV GitHub repository.
Apache-2.0 applies only to OMIV-authored files in the evidence repository. Upstream model and artifact terms remain applicable and unchanged; no upstream content is relicensed.