Upload GroundFlow training logs (4B/8B/32B) + refresh main results table

training_logs/32b_run1_a0.667_t1.5_iwa/runs/May05_03-21-17_n124-112-141/events.out.tfevents.1777951944.n124-112-141.8707.0

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:205e4266aa551483c0113a48ed3f0a31761bae5ce1ebf502f11cc17704086c43
+size 260236

training_logs/README.md

@@ -0,0 +1,25 @@
+# GroundFlow Training Logs
+
+Training logs for the **Confluent + IWA** method (paper: *Confluent Distillation for Vision-Language Models*) at three model scales. Each run uses `loss = α · L_roi + (1 − α) · L_KL` plus IWA (Importance-Weighted Attention) with γ on the twig branch.
+
+| Run | Base model | α | τ | IWA γ_init | LR | Steps | Final loss | Eval anchor |
+|-----|------------|---|---|------------|----|-------|------------|-------------|
+| `4b_run7b_a0.667_iwa` | Qwen3-VL-4B-Instruct | 0.667 | (run7b) | 1.0 | 5e-5 | — | see file | OCRBench=852 |
+| `8b_a0.667_t1.5_iwa_g1.0` | Qwen3-VL-8B-Instruct | 0.667 | 1.5 | 1.0 | 5e-5 | — | see file | OCRBench=866 |
+| `32b_run1_a0.667_t1.5_iwa` | Qwen3-VL-32B-Instruct | 0.667 | 1.5 | 1.0 | 5e-5 | 1173 | 0.4185 | (eval pending) |
+
+Each run directory holds:
+- `trainer_state.json` — full step-by-step log (loss, lr, grad_norm, epoch).
+- `runs/<TIMESTAMP>_<HOST>/events.out.tfevents.*` — TensorBoard scalars (32B run only).
+
+To plot from `trainer_state.json`:
+
+```python
+import json, matplotlib.pyplot as plt
+log = json.load(open("training_logs/32b_run1_a0.667_t1.5_iwa/trainer_state.json"))["log_history"]
+xs = [e["step"] for e in log if "loss" in e]
+ys = [e["loss"] for e in log if "loss" in e]
+plt.plot(xs, ys); plt.xlabel("step"); plt.ylabel("train loss"); plt.show()
+```
+
+Aggregated benchmark scores live in `doc/full_results_table.md`.