Upload PROJECT_RESUME_MANIFEST.md with huggingface_hub

PROJECT_RESUME_MANIFEST.md

@@ -0,0 +1,131 @@
+# DNAThinker Project Resume Manifest (2026-05-07)
+
+This document maps every artifact you need to **resume the DNAThinker
+paper work on a fresh machine**. Everything except local secrets is
+mirrored to public HuggingFace repos.
+
+---
+
+## 1. Code (GitHub)
+
+```bash
+git clone git@github.com:explcre/biomodel_reasoning_calling_study2.git
+git checkout mllm-integrate-server3   # active server3 branch
+# or `mllm-integrate` for the lab-merged trunk
+```
+
+- Latest server3 commit at snapshot time: `bc3c9aa` (paper at 54 pages)
+- Lab branch: `origin/mllm-integrate` (lab-cluster commits + my server3 merges)
+
+## 2. Model checkpoints on HuggingFace
+
+| Repo | What | Size |
+|---|---|---|
+| `explcre/dnathinker_t2_dual_xa` | T2 dual-XA Variant C (FT-100M+combined) | 992 MB |
+| `explcre/dnathinker_t3_trunks` | T3 edit-tight 500-bp Path-A production trunk | 507 MB |
+| `explcre/ntv3_rich_cond_mdlm_phase4_family` | T1 trunks (v3 paper-headline + v5full15m + others) | ~7 GB |
+| `explcre/phase7_multitask` | Joint MoE LoRAs (T1+T2+T3 r=64) | ~150 MB |
+| `explcre/phase8_rl` | All Phase-8 RL ckpts (incl.\ multi-seed reasoning RL, FIXED grid, KD/BIGLR) | ~5 GB |
+| `explcre/phase5_stage_a` / `phase5_stage_b` | Phase-5 SFT bases | varies |
+| `explcre/phase6_grpo` | Phase-6 GRPO RL | varies |
+
+The HF auto-uploader (`scripts/innovations/hf_auto_uploader.py`) keeps
+all of the above mirrored every 30 min.
+
+## 3. Paper-grade results (under `explcre/phase8_rl/_paper_results/`)
+
+- `tab_t1_v7predictor_n4000_bootstrap_ci.md` — FID/embed-cos CI on n=4000
+- `tab_t1_production_n200_bootstrap_ci.md` — FID 2.87 [1.65, 4.82] on n=200
+- `tab_rqrl_t1_bootstrap_ci.md` — TFG 0.4384 [0.351, 0.527] B=5000
+- `reasoning_rl_multiseed_summary.{md,json}` — T1/T2/T3 multi-seed aggregator
+- `cycle_70zz44_*` — production-route MoE end-to-end eval evidence
+- `cycle_70zz33_lraxis_bootstrap_ci.md` — T3 RL lr-axis bootstrap
+
+## 4. Multi-seed reasoning RL (under `explcre/phase8_rl/_reasoning_rl_multiseed/`)
+
+Each `exp_phase8_reasoning_grounded_rl_<task>_r128_alpha1_s<seed>_*`
+directory has `best.pt`, `log.jsonl`, `manifest.json`, and the
+matching `eval_reasoning_<task>_v7r128_postRL_alpha1_s<seed>_*/`
+has `score.json` + `score.md`.
+
+Tasks/seeds covered: T1 s=2,3 (cycle 70zz46); T3 s=2,3 (cycle 70zz48 parallel);
+T2 s=2 (cycle 70zz47). T2 s=3 was in flight at snapshot time.
+
+## 5. Claude memory (under `explcre/phase8_rl/_claude_memory/`)
+
+8 files covering user role, feedback rules, project context. Place
+under `~/.claude/projects/-workspace/memory/` on the new machine.
+
+## 6. Lab-cluster artifacts (lab side, under `explcre/phase7_multitask`)
+
+The lab cluster (3090×6, A6000×8, H100×4) ran SLURM jobs for the
+FULL_AUDIT 4-algo×3-seed grid + L7 edit-tight on 650M trunk.
+Outputs already merged into `mllm-integrate` and figures committed
+to `paper/figures/`.
+
+## 7. Public reference data (NOT in our HF; download fresh)
+
+- PsychENCODE source: `https://psychencode.synapse.org/`
+- NTv3 100M-post / 650M snapshots: `https://huggingface.co/InstaDeep/NTv3-...`
+- Qwen3.5-0.8B base: `https://huggingface.co/Qwen/Qwen2.5-0.5B` (or local)
+- JASPAR motif PWMs: `https://jaspar.genereg.net/`
+
+---
+
+## 🔒 What you must save LOCALLY (do NOT upload)
+
+Save these to your local machine — they are credentials and not
+publishable:
+
+| File | Where | Why |
+|---|---|---|
+| `/workspace/dnathinker/.env` | rsync to `~/dnathinker.env.bak` | OPENROUTER_API_KEY_{1..N} for reasoning expansion |
+| `~/.huggingface/token` (or `HF_TOKEN` env) | already on your machine | HF push permissions |
+| `~/.ssh/id_*` | already on your machine | GitHub push |
+| `~/.netrc` (if it exists) | already on your machine | git auth fallback |
+| `~/.kaggle/` (if used) | already on your machine | Kaggle data |
+| Any `*.env` under `/workspace/biomodel_reasoning_calling_study2/` | rsync | task-local secrets |
+
+```bash
+# Suggested local backup commands
+rsync -av root@<server3-host>:/workspace/dnathinker/.env ~/dnathinker.env.bak
+# (HF token already in ~/.huggingface/ on your local machine)
+```
+
+---
+
+## How to resume work on a fresh machine
+
+```bash
+# 1. Clone repo
+git clone git@github.com:explcre/biomodel_reasoning_calling_study2.git
+cd biomodel_reasoning_calling_study2/regureasoner_loop
+git checkout mllm-integrate-server3
+
+# 2. Install dependencies
+pip install -r requirements.txt   # if exists, else use existing env
+# Required: torch, transformers, peft, huggingface_hub, datasets,
+#          numpy, pandas, matplotlib, pyyaml, etc.
+
+# 3. Restore Claude memory (if using Claude Code)
+mkdir -p ~/.claude/projects/-workspace/memory/
+huggingface-cli download explcre/phase8_rl _claude_memory \
+  --local-dir ~/.claude/projects/-workspace/memory/ --include "_claude_memory/*"
+mv ~/.claude/projects/-workspace/memory/_claude_memory/* \
+   ~/.claude/projects/-workspace/memory/
+
+# 4. Restore .env (from your local backup)
+cp ~/dnathinker.env.bak /workspace/dnathinker/.env
+
+# 5. Pull critical model ckpts as needed
+# T3 edit-tight production trunk:
+mkdir -p /workspace/dnathinker/runs/
+huggingface-cli download explcre/dnathinker_t3_trunks \
+  exp_t3_edit_tight_20260505 --local-dir /workspace/dnathinker/runs/exp_t3_edit_tight_20260505
+
+# 6. Compile paper to verify
+cd paper && pdflatex main.tex && bibtex main && pdflatex main.tex && pdflatex main.tex
+```
+
+Done. Everything paper-grade is mirrored; nothing project-essential
+is local-only.