dummy-selfie-adapter-llama-3.2-1b

Dummy SelfIE adapter (random weights, Llama-3.2-1B-Instruct)

[!WARNING] This is not a trained model. Every weight in this repository is random. It exists only as a test fixture. Any output produced with it is noise by construction. Do not use it for inference, evaluation, or any result you intend to report.

What this is

A random-weight SelfIE adapter checkpoint, in the real selfie_adapters on-disk format, sized for meta-llama/Llama-3.2-1B-Instruct (hidden_dim = 2048).

Its only purpose is to let an end-to-end test of a SelfIE pipeline run on a laptop-sized GPU. The published, trained SelfIE adapters are 4096-wide and load only against the 8B base model, so nothing at 1B scale can load one. Without a fixture of the right width, a small-model test has to fabricate weights at run time — which means the test runs a code path the real run does not, and therefore proves less than it appears to.

What it proves, and what it does not

Does prove: the adapter loader, the safetensors metadata header, the dimension check, the projection math, and everything downstream of them (injection, generation, scoring, aggregation, file output) all run without crashing and produce the expected shapes.

Does not prove: anything at all about whether a SelfIE adapter recovers information from a hidden state. This adapter is untrained. It has learned nothing and cannot reveal anything. A green test says the plumbing works, not that the method works.

Contents

Architecture matches the real trained checkpoints (scalar_affine, normalize_input = true); only the weights differ. init_scale is set to the median L2 norm of the base model's input embedding rows, so the soft token lands at a plausible embedding scale for this model. A soft token far outside embedding scale makes generation degenerate for reasons unrelated to the shapes this fixture exists to test.

Provenance

Generated at seed 0 on CUDA, by make_smoke_weights.py in the project that publishes this fixture. init_scale = 0.9332589507102966, also recorded in this checkpoint's safetensors metadata header.

Verify by content, not by hash. Regenerating at seed 0 reproduces the tensors exactly, but not the file bytes: safetensors serializes the __metadata__ map in a per-process order, so two runs on one machine already differ by hash while holding identical tensors and identical metadata values.

init_scale is the one value that depends on the device. It is a median over the base model's embedding rows, and that reduction differs in its last bits (0.9332588315010071 on CPU). Each value is stable on its own device. bias is scaled by init_scale, so it shifts with it. The relative difference is ~1e-7 and irrelevant to a fixture whose purpose is shape checking.