diegoquinteiro/Gemma-3-270M-Observable

Gemma 3 270M Observable

Browser-oriented ONNX export of google/gemma-3-270m for the interactive laboratory in the Engenharia Assistida por IA course. The export uses the public unsloth/gemma-3-270m mirror of the checkpoint.

model_observable_f16_v2.onnx returns final next-token logits and the full attention matrix for each of the 18 transformer layers. It does not include a tuned lens or intermediate vocabulary predictions.

Inputs

• input_ids: int64 tensor shaped [batch, sequence];
• attention_mask: int64 tensor shaped [batch, sequence].

The graph returns next_token_logits and attention_01 through attention_18. Each attention tensor has shape [batch, 4, query sequence, key sequence].

Compression

Transformer weights use float16. Gemma's tied input and output embedding table uses row-wise int8. Public inputs and outputs remain float32 or int64.

For the validation prompt in observable-model.json, the browser artifact preserved all 10 highest-ranked next-token candidates from the original checkpoint. The maximum absolute attention error against the original checkpoint was 0.0602. Export-only error after quantization was 0.00163. This artifact is intended for teaching and inspection, not evaluation or production inference.

Limitations

• Gemma 3 270M is a small base model, not an instruction-following model.
• The graph exposes final next-token predictions only. It has no tuned lens for intermediate predictions.
• Attention weights describe values computed inside the model. They do not prove that a source token caused a prediction.