README.md

metadata

license: apache-2.0
base_model: EleutherAI/pythia-160m-deduped
library_name: onnx
pipeline_tag: text-generation
tags:
  - transformers.js
  - onnx
  - interpretability
  - attention
  - tuned-lens

Pythia-160M Observable

Browser-oriented ONNX export of EleutherAI/pythia-160m-deduped for the interactive laboratory in the Engenharia Assistida por IA course.

The repository contains two runtime artifacts:

• model_observable_q4.onnx, which returns final next-token logits, 13 residual-stream states, the full attention matrix and the selected query token's attention update for each of the 12 transformer layers;
• model_tuned_lens_q8.onnx, which applies the layer-specific translators from the pretrained AlignmentResearch/tuned-lens artifact and decodes an intermediate state into vocabulary logits.

Observable model inputs

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

The graph returns next_token_logits, hidden_state_00 through hidden_state_12, attention_01 through attention_12, and attention_output_01 through attention_output_12. Each attention tensor has shape [batch, heads, query sequence, key sequence].

Tuned lens inputs

• hidden_states: float32 tensor shaped [states, 768];
• layer_index: int64 tensor shaped [states], with values from 1 through 12.

Layer 12 uses the model's final normalization and unembedding without a translator. Earlier layers use the pretrained translator for that residual-stream position.

Quantization

Transformer matrix multiplications use weight-only Q4. The separate input and output embedding matrices use row-wise int8. The tuned lens translators and unembedding also use row-wise int8. Activations and public outputs remain float32.

The export manifest records the numerical checks performed against the PyTorch model and the original tuned lens. Quantization can change close-ranking tokens, so this artifact is intended for teaching and inspection, not evaluation or production inference.

Limitations

• Pythia-160M is a small base model, not an instruction-following model.
• Attention weights describe values computed inside the model. They do not prove that a source token caused a prediction.
• Intermediate predictions are diagnostic readouts from a tuned lens, not text generated by stopping the original model early.