Upload folder using huggingface_hub

.claude/memory/MEMORY.md

@@ -0,0 +1,2 @@
+- [Lotus PoC project context](project_lotus_poc.md) — Digital twin for Lotus's grocery, current state and what's next
+- [Background process patience](feedback_background_processes.md) — Don't check/kill background processes too early

.claude/memory/feedback_background_processes.md

@@ -0,0 +1,11 @@
+---
+name: Background process patience
+description: Don't check or kill background processes too aggressively — give them time to complete
+type: feedback
+---
+
+Don't check background processes every 2-3 minutes and don't kill them when output appears empty. Processes that load large models (BGE-M3, Qwen3-8B) take several minutes before producing output. Checking too early and killing leads to wasted GPU time and repeated restarts.
+
+**Why:** Multiple times during Lotus PoC, processes were killed while they were actually working fine — just hadn't produced output yet due to model loading or output buffering.
+
+**How to apply:** Set longer check intervals (10-15 min minimum for model-heavy tasks). Use `ps aux` to verify a process is actually dead before restarting, rather than assuming from empty output.

.claude/memory/feedback_no_summarize.md

@@ -0,0 +1,11 @@
+---
+name: Don't summarize raw data
+description: User wants raw outputs written to files, not summarized in chat
+type: feedback
+---
+
+When user asks for model outputs or examples, write the raw data to a file rather than summarizing or reformatting it in chat. The user wants to inspect the actual data themselves.
+
+**Why:** User explicitly said "i'm no need you to summarize, write the raw prompt+model_response into the file please"
+
+**How to apply:** When showing model outputs, prompts, or comparison data — dump to a JSON/file first, then tell the user where it is. Only provide brief stats (counts, file path) in chat.

.claude/memory/project_lotus_poc.md

@@ -0,0 +1,35 @@
+---
+name: Lotus PoC project context
+description: Digital twin for Lotus's Thailand grocery — current state, completed work, what's next
+type: project
+---
+
+## Project: Digital Twin for Lotus's (Thailand grocery/hypermarket)
+
+**Approach:** Li, Wei & Wang 2025 framework (Fine-Tuning + RAG) on Qwen3-8B
+
+### Completed:
+- Data prep: 60 users selected, 80/20 per-user chronological split
+- Format A: Shopping DNA Profiles (system prompts)
+- Format B: Decision point pairs (B1 purchase YES, B2 purchase NO 3:1 ratio with hard+easy negatives, B3 basket next-item, B4 promotion response)
+- RAG KB: 3,850 products in ChromaDB with BGE-M3 embeddings (1024 dim)
+- Training: 10 per-user LoRA adapters (users with 300+ pairs) — trained WITHOUT RAG in prompts
+- Evaluation (no RAG): Base 72.5% / LoRA 66% / GPT-5.4 77.5% purchase accuracy on 10 users
+
+### In progress:
+- Adding RAG context to training + inference (rebuild Format B with RAG, retrain, re-eval)
+- Gemini model name was wrong (gemini/gemini-3.1-pro not found)
+
+### Key findings so far:
+- LoRA needs 300+ training pairs to beat base (confirmed from Amazon PoC)
+- 3:1 B1:B2 ratio fixed the over-rejection problem (recall 65% → 86%)
+- RAG was built but never injected into prompts — fixing now
+- B4 test set needed negative examples (balanced now)
+
+### Data fields:
+- total_net_spend_amt = true line total (not net_spend_amt)
+- product_qty = units per txn-item line
+- retail_price = unit price before discount
+
+### Conda env: lotus-dt (Python 3.10)
+### BytePlus judge: DeepSeek V3 via litellm