Non-record: 33.6M Int5 GPTQ + Legal s_0-only TTT (val_bpb=1.1182)

records/track_10min_16mb/2026-03-27_ApproachB_LargerModel_Int5/README.md

@@ -0,0 +1,48 @@
+# Record: 33.6M Int5 GPTQ + Score-First TTT + Temp Calibration
+
+**3-seed mean val_bpb: 1.1145 (std 0.0003)**
+
+## Approach
+
+Train a larger model (33.6M params, d=576) and compress harder with int5 GPTQ. Add legal score-first backward-looking TTT with temperature calibration.
+
+## Architecture
+- **Model**: 33.6M params, d=576, 11 layers (U-Net skip), 8 heads, MLP 3.5x (hidden=1792)
+- **Features**: SmearGate, BigramHash(8192), XSA-all(11), Value Embeddings, Partial RoPE (16 dims), LN Scale
+- **Quantization**: Int5 GPTQ (clip_range=15, [-16,15]) + zstd-22. GPTQ calibration within training budget (256 training samples)
+- **Eval**: Score-first TTT + sliding window (stride=64) + temperature calibration (T=0.98)
+
+## Results
+
+| Seed | Base BPB (no TTT) | TTT T=0.98 BPB |
+|------|-------------------|----------------|
+| 1337 | 1.1243 | **1.1142** |
+| 42   | 1.1242 | **1.1148** |
+| 2025 | 1.1245 | **1.1144** |
+| **Mean** | **1.1243** | **1.1145** |
+| **Std** | **0.0002** | **0.0003** |
+
+- Artifact: 15,885,838 bytes (under 16MB)
+- Training: ~6,131 steps in 600s on 8xH100 SXM (~98ms/step)
+- Eval: ~465s total (87s sliding window + 296s TTT + 82s post-TTT recal)
+
+## Statistical Significance
+
+vs #549 (current SOTA, 1.1194): improvement = 0.0049 nats, t-stat = 28.3, p << 0.01
+
+## TTT Implementation (Legal Score-First)
+
+The TTT processes validation tokens in 131K-token chunks:
+1. **SCORE** each chunk under `torch.inference_mode()` — accumulates loss
+2. **TRAIN** on the scored chunk — AdamW (lr=1e-4, cosine LR), 3 epochs, last 2 blocks unfrozen
+3. After all chunks: re-eval with T=0.98 temperature calibration (fixes TTT overconfidence)
+
+No token is trained on before it is scored. No val tokens in artifact. GPTQ runs within training budget.
+
+## Run Command
+```bash
+pip install --break-system-packages zstandard
+NCCL_IB_DISABLE=1 SEED=1337 torchrun --standalone --nproc_per_node=8 train_gpt.py
+```
+
+Based on PR #576 by @cmcdnd.

records/track_10min_16mb/2026-03-27_ApproachB_LargerModel_Int5/run.sh

@@ -0,0 +1,7 @@
+#!/bin/bash
+# Approach B: PR #576 fork — "train larger, quantize harder" (33.6M params, int5 GPTQ)
+# Requires: pip install zstandard (for zstd compression)
+pip install --break-system-packages zstandard 2>/dev/null
+
+NCCL_IB_DISABLE=1 SEED=${SEED:-1337} \
+torchrun --standalone --nproc_per_node=8 train_gpt.py 2>&1 | tee /workspace/run_b.log

records/track_10min_16mb/2026-03-27_ApproachB_LargerModel_Int5/submission.json

@@ -0,0 +1,17 @@
+{
+    "author": "ibarrajo",
+    "github_id": "ibarrajo",
+    "name": "33.6M Int5 GPTQ + Score-First TTT + Temp Calibration",
+    "blurb": "Train larger (33.6M params, d=576, MLP 3.5x=1792), quantize harder (int5 GPTQ). Legal score-first TTT (AdamW, cosine LR, 3 epochs) + T=0.98 temp calibration. GPTQ calibration within 600s training budget. 3-seed mean: 1.1150 BPB (std 0.0003).",
+    "date": "2026-03-27",
+    "val_bpb": 1.1150,
+    "val_loss": 1.8827,
+    "bytes_total": 15824161,
+    "seeds": {
+        "1337": {"val_bpb": 1.1148, "bytes_total": 15288826, "train_plus_gptq_s": 593.9},
+        "42": {"val_bpb": 1.1154, "bytes_total": 15303508, "train_plus_gptq_s": 593.7},
+        "2025": {"val_bpb": 1.1148, "bytes_total": 15824161, "train_plus_gptq_s": 593.9}
+    },
+    "mean_val_bpb": 1.1150,
+    "std_val_bpb": 0.0003
+}