docs: ROADMAP_SCRATCH Step 0-10 with status

docs/PROGRESS.md

@@ -29,6 +29,7 @@ Har kaam ke baad yaha entry karo taaki yaad rahe kab kya kiya tha.
 - [x] 2026-09-17: Kaggle smoke FAIL (`data/raw` HF me excluded) -> `train.py` smoke fallback (built-in dummy lines), Kaggle re-test PASS (24 blocks, loss 5.38->5.37), `docs/TEST_REPORT.md` banaya, HF push OK
 - [x] 2026-09-17: full code-phase suite 5/5 PASS (params 241.7M/166.2M, sliding unit, toy router balanced max 0.203 + loss 6.18->5.95, tok fertility 1.38 unk 0, train smoke), TEST_REPORT update, HF push
 - [x] 2026-09-17: POLICY LOCK — `docs/TESTING_POLICY.md` (saare tests Kaggle pe, local pe koi test nahi), TEST_REPORT header update, HF push
+- [x] 2026-09-17: `docs/ROADMAP_SCRATCH.md` banaya (Step 0-10, status: 0-2 done, agla Step 3 toy-Kaggle), HF push
 - [ ] TODO: toy router-balance test (Kaggle), FINAL tokenizer (500k lines wait), dense baseline
 
 ## Kaise note likhna hai

docs/ROADMAP_SCRATCH.md

@@ -0,0 +1,50 @@
+# ROADMAP — Model Scratch se Deployment tak (ViuMini-MoE-242M status ke saath)
+> Har step ka status: [x] done / [ ] bacha. Evidence file bracket me.
+> Update rule: step complete hote hi tick + PROGRESS + TEST_REPORT entry + HF push.
+
+## Step 0 — Goal + design lock [x] DONE
+- Model type, size, language mix fix (Hinglish 40:30:30, 28L MoE 241.7M/166.2M LOCK)
+- Evidence: `docs/VIU1_MOE_PLAN.md`, `model/configs/viu1_moe_config.yaml`
+
+## Step 1 — Repo + code skeleton [x] DONE
+- Folder structure, arch code (`viu1_moe.py`), config, requirements, HF card
+- Evidence: `README.md`, `model/scripts/viu1_moe.py`, HF https://huggingface.co/ViuAI/ViuMini-MoE-242M
+
+## Step 2 — Audit + bugfix [x] DONE
+- Sliding-mask, MoE dispatch, aux-loss, batch, ckpt, paths — 10 critical fix verify
+- Evidence: `docs/TEST_REPORT.md` (5/5 suite PASS)
+
+## Step 3 — Toy test (Kaggle) [ ] BACHA
+- 2L tiny MoE pe router-balance (max expert <60%), balance-loss lock, switch-gate on/off
+- Success: usage report + loss-down curve TEST_REPORT me
+- Fail rule: yaha fail = design badlo, aage mat badho
+
+## Step 4 — Real data collect [ ] BACHA (sabse important)
+- 4-5B tokens 40:30:30 (2B Hinglish + 1.5B Hindi + 1.5B English) + 200k CoT Hinglish
+- Clean + dedup + eval split, 0 leakage
+- Evidence: `data/processed/` + dedup report + `ViuAI/viu-mini-pretrain-40-30-30` dataset repo
+
+## Step 5 — FINAL tokenizer [ ] BACHA
+- 500k lines pe 48k BPE train, fertility Hinglish ≤1.5 / Hindi ≤1.8 / unk=0
+- Evidence: `tokenizer/outputs/tokenizer.json` (FINAL tag) + coverage log
+
+## Step 6 — Dense baseline (GPU) [ ] BACHA
+- MoE off karke ~135M dense train, loss smooth down + Hinglish fluent sample
+- Evidence: ckpt + val PPL + sample log (TEST_REPORT)
+
+## Step 7 — Full MoE pretrain (GPU 2-4 hafte) [ ] BACHA
+- 9 MoE layers on, BF16, ckpt har 5k, early-stop (expert >60% for 2k steps → roko)
+- Success: dense se +5% better + router balanced
+
+## Step 8 — Eval + inference [ ] BACHA
+- PPL (3 langs), CoT accuracy, human rating + `generate.py` (top-k/temp)
+- Evidence: eval numbers + samples PROGRESS me
+
+## Step 9 — SFT / alignment (optional) [ ] BACHA
+- Instruction + DPO data pe finetune (plan me baad me)
+
+## Step 10 — Release [ ] BACHA
+- FINAL ckpt + model card + public HF + version tag
+
+## Abhi kaha hai?
+Step 0-2 done. Agla: **Step 3 toy test (Kaggle)** → Step 4 data (tumhara kaam) → Step 5 tokenizer.