| { |
| "version": "v2", |
| "input": { |
| "shape": "[T, 512]", |
| "dtype": "float32", |
| "what": "EmbeddingGemma 2 audio soft tokens (embed_audio output), 25 per second" |
| }, |
| "context": "bidirectional within the window", |
| "architecture": { |
| "name": "a", |
| "kind": "transformer encoder from scratch", |
| "layers": 6, |
| "d_model": 512, |
| "heads": 8, |
| "ffn": 2048, |
| "activation": "gelu", |
| "norm": "pre-LayerNorm", |
| "positions": "rotary (base 10000)", |
| "input": "LayerNorm -> Linear 512->512", |
| "output": "LayerNorm -> Linear 512->768", |
| "params": 19573762 |
| }, |
| "output": { |
| "vector": "[T, 768] L2-normalized", |
| "blank_logit": "[T]" |
| }, |
| "logits": "scale * cos(vector, table[w]) for w in vocab; blank from blank_logit", |
| "scale": 216.3509979248047, |
| "vocab": "vocab.txt", |
| "table": { |
| "file": "word_table.npy", |
| "dtype": "float16", |
| "shape": "[200000, 768]", |
| "prompt": "task: search result | query: ", |
| "normalized": true |
| }, |
| "decode": "greedy argmax over N+1, collapse repeats, drop blank", |
| "window": { |
| "seconds": 30, |
| "left_overlap_tokens": 96, |
| "right_extra_tokens": 1, |
| "note": "encoder is causal; the 83 tokens at the start of a window and its last token differ from a continuous run (fixtures/long-speckled-band: soft_tokens_w30 vs soft_tokens_continuous)" |
| }, |
| "weights": { |
| "file": "reader.pt", |
| "format": "torch state_dict, float32; training matmuls in TF32 (A100), evaluation in fp32", |
| "code": "train/model.py Reader(arch, n_vocab, k, layers)", |
| "keys": "body.* (reader body), blank.weight [1, 768], blank.bias [1], log_scale (scale = exp(log_scale))" |
| }, |
| "n_vocab": 200000, |
| "wer": { |
| "dev-clean": { |
| "wer": 0.03754, |
| "utterances": 2703, |
| "ref_words": 54402 |
| }, |
| "dev-other": { |
| "wer": 0.10263, |
| "utterances": 2864, |
| "ref_words": 50948 |
| }, |
| "test-clean": { |
| "wer": 0.04207, |
| "utterances": 2620, |
| "ref_words": 52576 |
| }, |
| "test-other": { |
| "wer": 0.10422, |
| "utterances": 2939, |
| "ref_words": 52343 |
| }, |
| "fixtures": { |
| "wer": 0.03697, |
| "utterances": 20, |
| "ref_words": 541 |
| }, |
| "decode": "greedy, no language model" |
| } |
| } |
