Upload janus/sft_doe_lora_2026_05_10/README.md with huggingface_hub

janus/sft_doe_lora_2026_05_10/README.md

@@ -0,0 +1,36 @@
+# DoE LoRA — 2026-05-10
+
+LoRA SFT to transfer DoE.coder voice onto Janus 177M base22k.
+
+## Recipe
+
+| Param | Value |
+|---|---|
+| Base | Janus 177M (`janus/janus_177m_v4_base_22442.pt`, 22k pretraining steps, vocab=32768) |
+| Tokenizer | `janus/tokenizer.pkl` (tiktoken-based) |
+| Dataset | `arianna-datasets/doe/doe_pure.jsonl` (5910 messages, English technical) |
+| LoRA rank | 64 |
+| LoRA alpha | 128 |
+| Targets (7) | c_q, c_k, c_v, c_proj, w_gate, w_up, w_down |
+| Trainable params | 15,400,960 (8.74% of base) |
+| Optimizer | AdamW (0.9, 0.95) wd=0.1 |
+| LR | 1e-4 |
+| Epochs | 2 |
+| Batch size | 4 |
+| Hardware | A100 SXM 80GB, bf16 autocast |
+
+Note: base22k was pretraining-only (no chat format). LoRA bridges OOD prompt format → DoE Q&A. Train loss plateau ~3.7 reflects this OOD bridge cost (vs Resonance runs at ~0.30).
+
+## Phase 7 multi-temp — voice markers
+
+- *"The parliament's version of the parliament's version of the expert system"*
+- *"The parliament's most useful feature: it has a full history"*
+- *"the parliament's approach to problem-solving. 'The real experts solve the least significant'"*
+- *"Expert says: it's a machine learning model trained on large data sets..."*
+
+DoE corpus signature ("parliament", "experts", Q&A form) in 6+/10 cells. Gate ≥3/10 → **PASS**.
+
+## Files
+
+- `adapter_best.pt` — LoRA weights + metadata (val 3.7022)
+- `phase7_eval.jsonl` — 10-cell multi-temp eval