Add accepted real2 targeted logic/math SFT candidate

real2_targeted_logic_math_sft_step040/README.md

@@ -0,0 +1,20 @@
+---
+license: apache-2.0
+language:
+- en
+tags:
+- causal-lm
+- sft
+- logic-reasoning
+---
+
+# TestGeniy real2 targeted logic/math SFT
+
+Experimental best-measured candidate derived from TestGeniy dialogue SFT v6. Only layers.23 and final_norm were trained on original train splits from OASST1, HelpSteer2, FOLIO, ARC, StrategyQA, GSM8K, and Hendrycks MATH. Benchmark test/validation rows, model-generated answers, and MTP were excluded.
+
+| checkpoint | GSM8K | MATH-500 | ARC-Challenge | FOLIO | composite |
+|---|---:|---:|---:|---:|---:|
+| v6 | 24/100 | 7/100 | 26/100 | 29/100 | 21.5 |
+| candidate | 23/100 | 7/100 | 29/100 | 35/100 | 23.5 |
+
+The candidate is not silently promoted over v6: GSM8K is one point lower, while logic benchmarks improve and MATH is unchanged. Balanced human preference accuracy on 80 OASST1/HelpSteer2 pairs is unchanged at 55%.

real2_targeted_logic_math_sft_step040/benchmark_comparison.json

@@ -0,0 +1,20 @@
+{
+  "v6_fixed100": {
+    "gsm8k": 24.0,
+    "math500": 7.0,
+    "arc_challenge": 26.0,
+    "folio": 29.0,
+    "composite": 21.5
+  },
+  "candidate_fixed100": {
+    "gsm8k": 23.0,
+    "math500": 7.0,
+    "arc_challenge": 29.0,
+    "folio": 35.0,
+    "composite": 23.5
+  },
+  "human_preference_balanced80": {
+    "v6": 55.0,
+    "candidate": 55.0
+  }
+}

real2_targeted_logic_math_sft_step040/manifest.json

@@ -0,0 +1,83 @@
+{
+  "checkpoint": "/marimo/testgeniy_real2/checkpoints/targeted_logic_math_sft_step040.pt",
+  "training": {
+    "algorithm": "frozen-backbone targeted SFT",
+    "trainable_scope": "layers.23 + final_norm",
+    "steps": 40,
+    "micro_batch": 2,
+    "grad_accum": 4,
+    "lr": 1e-07,
+    "max_len": 768,
+    "seed": 20260819,
+    "real_only": true,
+    "synthetic": false,
+    "mtp": false,
+    "source_counts": {
+      "logic": 186,
+      "math": 134
+    },
+    "logic_rows": 4409,
+    "math_rows": 14972,
+    "sft_fingerprint": "51c467c9c37c19bc99881e35e5db15d4780004d94960bfb15854efd01695b154"
+  },
+  "history": [
+    {
+      "step": 1,
+      "loss": 3.69091,
+      "grad_norm": 4.0295,
+      "lr": 1.25e-08,
+      "last_domain": "logic"
+    },
+    {
+      "step": 10,
+      "loss": 3.51586,
+      "grad_norm": 4.8938,
+      "lr": 1e-07,
+      "last_domain": "math"
+    },
+    {
+      "step": 20,
+      "loss": 3.16587,
+      "grad_norm": 2.0741,
+      "lr": 1e-07,
+      "last_domain": "math"
+    },
+    {
+      "step": 30,
+      "loss": 3.50732,
+      "grad_norm": 3.7319,
+      "lr": 1e-07,
+      "last_domain": "logic"
+    },
+    {
+      "step": 40,
+      "loss": 3.27809,
+      "grad_norm": 3.7667,
+      "lr": 1e-07,
+      "last_domain": "math"
+    }
+  ],
+  "accepted": true,
+  "acceptance_rule": "fixed suite composite improves without any protected-domain regression",
+  "acceptance_reason": "fixed100 composite 23.5 vs v6 21.5; GSM8K 23 vs 24, MATH500 7 vs 7, ARC 29 vs 26, FOLIO 35 vs 29; balanced human preference 55 vs 55",
+  "benchmark_comparison": {
+    "v6_fixed100": {
+      "gsm8k": 24.0,
+      "math500": 7.0,
+      "arc_challenge": 26.0,
+      "folio": 29.0,
+      "composite": 21.5
+    },
+    "candidate_fixed100": {
+      "gsm8k": 23.0,
+      "math500": 7.0,
+      "arc_challenge": 29.0,
+      "folio": 35.0,
+      "composite": 23.5
+    },
+    "human_preference_balanced80": {
+      "v6": 55.0,
+      "candidate": 55.0
+    }
+  }
+}

real2_targeted_logic_math_sft_step040/targeted_logic_math_sft_step040.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:08a7c5e74c0445ea772da7951e7192920cb167f68fe02d2f07157bc061e9cda7
+size 2203292769