ViuAI commited on
Commit
bfaccbf
Β·
verified Β·
1 Parent(s): b2d2e61

step3: fair-50 matrix D+E, lock logic

Browse files
docs/PROGRESS.md CHANGED
@@ -35,6 +35,7 @@ Har kaam ke baad yaha entry karo taaki yaad rahe kab kya kiya tha.
35
  - [x] 2026-09-17: improve approved β€” toy me Config C (0.05, 50 steps) joda, HF push β€” Kaggle re-run pending
36
  - [x] 2026-09-17: Step 3 Config C Kaggle PASS (max 0.199, sab experts β‰₯5.5%, loss 4.55 β€” A/B se best), TEST_REPORT entry, HF push β€” 0.05 lock approval pending
37
  - [x] 2026-09-17: lock DEFERRED (user: best result aane tak lock nahi) β€” config 0.01 unchanged, Step 3 open, HF push
 
38
  - [x] 2026-09-17: `docs/ROADMAP_SCRATCH.md` banaya (Step 0-10, status: 0-2 done, agla Step 3 toy-Kaggle), HF push
39
  - [ ] TODO: toy router-balance test (Kaggle), FINAL tokenizer (500k lines wait), dense baseline
40
 
 
35
  - [x] 2026-09-17: improve approved β€” toy me Config C (0.05, 50 steps) joda, HF push β€” Kaggle re-run pending
36
  - [x] 2026-09-17: Step 3 Config C Kaggle PASS (max 0.199, sab experts β‰₯5.5%, loss 4.55 β€” A/B se best), TEST_REPORT entry, HF push β€” 0.05 lock approval pending
37
  - [x] 2026-09-17: lock DEFERRED (user: best result aane tak lock nahi) β€” config 0.01 unchanged, Step 3 open, HF push
38
+ - [x] 2026-09-17: best-path approved β€” toy me D (0.01/50) + E (0.02/50) jode, fair-50 lock logic, HF push β€” Kaggle re-run pending
39
  - [x] 2026-09-17: `docs/ROADMAP_SCRATCH.md` banaya (Step 0-10, status: 0-2 done, agla Step 3 toy-Kaggle), HF push
40
  - [ ] TODO: toy router-balance test (Kaggle), FINAL tokenizer (500k lines wait), dense baseline
41
 
model/scripts/toy_router_test.py CHANGED
@@ -4,7 +4,9 @@ Toy router-balance test β€” Step 3 (Kaggle, CPU enough, ~5 min)
4
  A) balance_weight=0.01 (current lock candidate)
5
  B) balance_weight=0.02 (stronger balancing)
6
  - Har config: train steps on dummy Hinglish lines, phir expert usage report
7
- - Configs: A) bal 0.01 / 20 steps, B) bal 0.02 / 20 steps, C) bal 0.05 / 50 steps (improve attempt)
 
 
8
  - Verdict: max expert share <60% = PASS, warna FAIL (design badlo, aage mat badho)
9
  - Report: output ko copy karke docs/TEST_REPORT.md me paste karo
10
 
@@ -114,13 +116,20 @@ def main():
114
  ra = run_config("A", 0.01, tok, blocks, steps=20)
115
  rb = run_config("B", 0.02, tok, blocks, steps=20)
116
  rc = run_config("C", 0.05, tok, blocks, steps=50)
 
 
117
  print("\n--- COPY TO TEST_REPORT ---")
118
- for r in (ra, rb, rc):
119
  print(f"- Config {r['tag']} (bal {r['bal_w']}): loss {r['loss_first']}->{r['loss_last']}, "
120
  f"max_share {r['max_share']}, usage {r['usage']} => {r['verdict']}")
121
- locked = "A (0.01)" if ra["verdict"] == "PASS" and ra["max_share"] <= min(rb["max_share"], rc["max_share"]) else \
122
- "B (0.02)" if rb["verdict"] == "PASS" and rb["max_share"] <= rc["max_share"] else \
123
- "C (0.05)" if rc["verdict"] == "PASS" else "NONE β€” design badlo, aage mat badho"
 
 
 
 
 
124
  print(f"locked balance_weight: {locked}")
125
  print("--- END ---")
126
 
 
4
  A) balance_weight=0.01 (current lock candidate)
5
  B) balance_weight=0.02 (stronger balancing)
6
  - Har config: train steps on dummy Hinglish lines, phir expert usage report
7
+ - Configs: A) bal 0.01 / 20 steps, B) bal 0.02 / 20 steps (history),
8
+ C) bal 0.05 / 50 steps, D) bal 0.01 / 50 steps, E) bal 0.02 / 50 steps
9
+ - Fair lock sirf 50-step group (C/D/E) se β€” steps ka asar alag ho jayega
10
  - Verdict: max expert share <60% = PASS, warna FAIL (design badlo, aage mat badho)
11
  - Report: output ko copy karke docs/TEST_REPORT.md me paste karo
12
 
 
116
  ra = run_config("A", 0.01, tok, blocks, steps=20)
117
  rb = run_config("B", 0.02, tok, blocks, steps=20)
118
  rc = run_config("C", 0.05, tok, blocks, steps=50)
119
+ rd = run_config("D", 0.01, tok, blocks, steps=50)
120
+ re_ = run_config("E", 0.02, tok, blocks, steps=50)
121
  print("\n--- COPY TO TEST_REPORT ---")
122
+ for r in (ra, rb, rc, rd, re_):
123
  print(f"- Config {r['tag']} (bal {r['bal_w']}): loss {r['loss_first']}->{r['loss_last']}, "
124
  f"max_share {r['max_share']}, usage {r['usage']} => {r['verdict']}")
125
+ # fair lock: sirf 50-step group (C/D/E) compare karo
126
+ fair = [r for r in (rc, rd, re_) if r["verdict"] == "PASS"]
127
+ if fair:
128
+ best = min(fair, key=lambda r: r["max_share"])
129
+ tagmap = {"C": "C (0.05)", "D": "A/D-line (0.01)", "E": "B/E-line (0.02)"}
130
+ locked = tagmap[best["tag"]] + f" β€” fair-50 best max_share {best['max_share']}"
131
+ else:
132
+ locked = "NONE β€” design badlo, aage mat badho"
133
  print(f"locked balance_weight: {locked}")
134
  print("--- END ---")
135