Step 1700 ONNX, loss 1.97

README.md

@@ -42,21 +42,21 @@ ByteEmbed → Conv1D×3 → BiAttention×2 → Pool → Matryoshka Heads
 
 ## Performance
 
-Benchmarked on synthetic data (500 samples, 1024 bytes max, base tier, 800 training steps):
+Benchmarked on synthetic data (500 samples, 1024 bytes max, base tier, 1700 training steps):
 
 | Head | Accuracy | Support |
 |------|----------|---------|
 | coarse | 100.0% | 500 |
 | modality | 100.0% | 500 |
-| subtype | 98.4% | 128 |
-| code_lang | 54.2% | 48 |
-| text_lang | 88.6% | 35 |
+| subtype | 93.8% | 128 |
+| code_lang | 41.7% | 48 |
+| text_lang | 94.3% | 35 |
 | file_mime | 100.0% | 131 |
-| risk (mAP) | 99.6% | — |
+| risk (mAP) | 100.0% | — |
 
-- **Inference**: ~8ms per sample on CPU (ONNX Runtime)
+- **Inference**: ~13ms per sample on CPU (ONNX Runtime)
 - **Model size**: ~200KB (FP32 ONNX)
-- **Loss**: 2.72 eval_loss (step 800)
+- **Loss**: 1.97 eval_loss (best, step 1700)
 
 > **code_lang** accuracy (54.2%) reflects 62-class coverage; improves with longer sequences (>256 bytes). v0.2 will target better code language discrimination.
 

picotype_base.onnx

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:ee2d91719a48ccbef5dc38b11c951d29194928b41e42cf7667edc0cb83bb47db
+oid sha256:43cbde196ac4450202bb70d64a56eebab538f61f2d0a4eb5d24fc7499182f395
 size 210599

picotype_base.onnx.data

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:bcb1f8d9a325ee53b2e3a81bc4f340bd53033b045234cb2823433a8481b899fa
+oid sha256:4936334bde1484726130aa93a90ec4b9d840ae6a2c060969eb7cc307ceb535e9
 size 9045864

picotype_pro.onnx.data

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:91be2c53c0347bbb5f48b1b833dee344126d178db37b7fc7e6257dd947996c45
+oid sha256:1fdf41552e630a3717ab8ff0a2ffdd924c0b4a2ca6d80835378552c92f430ace
 size 9401856

picotype_small.onnx.data

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:e5a253604ee5bed7117d638a767d675eb596634b7507c3fc82b271243ae960df
+oid sha256:99f2f03e785bdacad8737c502c490e0cbf7deddb6de151b5743625bf0c753c97
 size 8926720

picotype_tiny.onnx.data

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:11a8e41e59a916193832bccc617cc661ef149640ea9c566f07b561bd31e1ea49
+oid sha256:5604aac5f7d42fc146bbf3463fded4340958a1e0c563422d1a2928f37f3cf401
 size 8882176