Upload walkthrough.md with huggingface_hub

walkthrough.md

@@ -194,7 +194,7 @@ python -c "import torch, numpy, safetensors, yaml; print('ok')"
 - **Eval results (1000 samples, base tier, 5.6ms)**:
   - coarse 100%, modality 100%, subtype **98.4%**, code_lang **53.9%**, text_lang **100%**, file_mime 100%, risk mAP 100%
   - code_lang improved **+9.9%** (43.96% → 53.85%) via real GitHub code samples
-- **ONNX export**: All 4 tiers re-exported from best.pt (~207-209 KB on disk with external data, opset 18).
+- **ONNX export**: All 4 tiers re-exported from best.pt (~9 MB each, self-contained single file, opset 18, `external_data=False`).
 - **HF Model**: `huggingface.co/eulogik/pico-type` — ONNX models at root level + checkpoints/ directory, model card, paper scaffold.
 - **HF Space**: `huggingface.co/spaces/eulogik/pico-type` — Gradio app **label lists fixed** (Jun 18 2026). Root cause: `gradio_app.py` had different label ordering than `labels.py` (text_lang had `"ar","hi"` instead of `"id","ms"`; file_mime was completely different set). Also fixed `np.bool_`→`bool` for NumPy 2.x compat (Space runs Python 3.13).
 - **PyPI**: `pico-type` v0.1.3 published at https://pypi.org/project/pico-type/0.1.3/.
@@ -341,10 +341,19 @@ python -c "import torch, numpy, safetensors, yaml; print('ok')"
 - **INT8**: ONNX shape inference fails — `[ShapeInferenceError] Inferred shape (192) vs (12)`. Multi-head architecture (shared 192-dim pooled vector → 7 linear layers with different output dims) confuses shape inference.
 - **FP16**: `onnxconverter_common.float16` succeeds but produces type mismatch errors in ONNX Runtime. Could not resolve with `op_block_list`.
 
-### Fresh training from scratch (in progress)
-- Started training from scratch with diverse generator + real GitHub data (30% ratio) + high coarse weight (8.0)
-- 4000 steps, batch_size=16, 15 samples/lang from 62 languages
-- Located at: `checkpoints/fresh/train.log`
+### Fresh training from scratch experiment (Jun 18, discarded)
+- Tried training from scratch with diverse generator + 647 real GitHub samples (30% ratio) + high coarse weight (8.0)
+- 4000 steps, batch_size=16, best eval loss 2.42
+- **Result**: Worse than fine-tuned model — real-world accuracy only 28.6% vs 52.4% from fine-tuned best.pt
+- **Lesson**: Training from scratch with high coarse weight over-prioritizes coarse classification at expense of code_lang/text_lang. Fine-tuning from a good synthetic checkpoint with gradual head-weight adjustments works better.
+- **Status**: Discarded. Production model remains `checkpoints/best.pt` (52.4% real-world accuracy).
+
+### v0.1.6: Docs, model card, HF fixes (Jun 18)
+- Uploaded `walkthrough.md`, `docs/PLAN.md`, `MODEL_CARD.md` to HF model repo (fixes broken links on HF model page)
+- Updated README.md with badges (PyPI, CI, DOI), eulogik branding, real-world accuracy metrics
+- Updated MODEL_CARD.md with full training/architecture details, citation, eulogik branding
+- Updated paper/main.tex with real-world eval, diverse generator details, higher coarse weight
+- Created HF org card content for eulogik organization page (paste at huggingface.co/eulogik)
 
 ---