Apr 8 00:13 Upload news summarization eval results for 69d59c104adb81dd2de75563 Apr 7 23:57 Upload news summarization eval results for 69d5986c4adb81dd2de75558 23:37 Fix HF Jobs result upload path expansion 23:29 Install runtime deps inside HF Jobs launcher 23:18 Document Hugging Face Jobs path for news summarization eval 23:18 Add Hugging Face Jobs launcher for news summarization eval 22:49 Narrow Colab snapshot download to news summarization files 22:47 Lazy-load shared providers for lighter notebook imports 22:37 Use lighter Colab defaults for news summarization notebook 22:37 Improve Colab runner progress visibility 21:38 Use valid default model in news summarization notebook 18:35 Include shared eval code in Colab snapshot download 18:27 Use HF snapshot download in news summarization notebook 18:19 Fix notebook result inspection cell 18:13 Link Colab notebook in news summarization README 18:13 Add Colab notebook for news summarization eval 16:46 Add news summarization eval harness Apr 5 16:20 Upload local-intelligence eval results for 69d289624adb81dd2de74caf 15:23 Upload local-intelligence eval results for 69d27d17074a09ff9be109c6 15:09 Upload local-intelligence eval results for 69d27a3e074a09ff9be109a9 04:55 Upload local-intelligence eval results for 69d1ea7b4adb81dd2de74b7a 04:52 Sync local-intelligence eval harness 04:40 Sync local-intelligence eval harness 04:07 Upload local-intelligence eval results for 69d1de21074a09ff9be10463 03:44 Sync local-intelligence eval harness 02:00 Upload local-intelligence eval results for 69d1c0104adb81dd2de74ade 01:51 Sync local-intelligence eval harness Apr 4 21:56 Sync local-intelligence eval harness 21:56 Sync local-intelligence eval harness 21:56 Sync local-intelligence eval harness Mar 7 17:54 🧪 add segment classifier, protocol processor test harness, and full chain tests - Part 4 blog: segment classifier synopsis with processor + chain test results
- Segment classifier: training script, eval harness, 50 labeled mixed dictations
- Protocol processor: updated with needs-llm detector, test harness (92 unit + 100 eval)
- Full chain test: 12 cases through classifier → model → reassembly
- Model weights: segment-classifier-model.json, needs-llm-model.json
- Eval data: 100 dictation-to-command pairs across 31 categories Mar 6 23:03 Fix repo_type for test data download 23:02 Add full eval notebook for Qwen3.5-2B adapter Mar 5 22:19 ✨ add Qwen3-1.7B notebook: text-only, no VL hacks, clean training path 21:51 💾 auto-save adapter immediately after training 20:39 🧠 disable thinking mode: empty <think> block + faster eval with eos_token_id 18:52 🐛 fix training cell: manual ChatML format, list return for Unsloth SFTTrainer 17:07 ⚡ skip install step if unsloth already loaded 16:54 ✨ add Qwen3.5-2B fine-tuning notebook for Colab Unsloth + LoRA on free T4. Full pipeline: load, train 3 epochs,
evaluate on 630 test cases, save/push adapter to HuggingFace. 16:29 📝 add Part 3 outline: the 40-millisecond classifier gate Trained embedding classifier (logistic regression on NLEmbedding)
decides whether dictation needs LLM normalization. 100% accuracy
on held-out data, trained in 40ms on 120 examples using BLAS.
Covers: hand-crafted vs trained classifier, Accelerate/BLAS speedup,
fan-out architecture, and the routing gate concept. Mar 4 23:13 🧪 initial commit — voice-to-syntax training lab Processor, pipeline, eval data, training infrastructure, and blog drafts
from the voice dictation → programming syntax experiments.
Architecture: LLM normalizer → procedural processor (split approach).
Zero-training pipeline achieves 47% on mixed-difficulty eval (2x baseline). 23:18 initial commit 
