Document the SpaMo training checkpoint

README.md

@@ -15,6 +15,7 @@ assets/            mt5_tokenizer/ flan_t5_tokenizer/ yolo/yolo11n.onnx mediapipe
 unisign/<name>/    projector.onnx encoder.onnx decoder.onnx prefix.npy meta.json
 spamo/shared/      clip_vit_s2.onnx(+.data) videomae.onnx(+.data) hand_clip.onnx(+.data)
 spamo/<name>/      projector.onnx(+.data) encoder.onnx(+.data) decoder.onnx(+.data) prompt_embeds.npy meta.json
+checkpoints/spamo/ spamo_pjm_ms_dual_bleu5.36.ckpt (PyTorch Lightning, not used by signsdk)
 ```
 
 ## Models (test-set BLEU-4, ✓ = uploaded)
@@ -41,3 +42,11 @@ spamo/<name>/      projector.onnx(+.data) encoder.onnx(+.data) decoder.onnx(+.da
 BLEU-4 = reported PJM test score. Decoding faithful to the original eval
 (Uni-Sign beam=4; SpaMo beam-sample num_beams=5, top_p=0.9). fp32 chosen for deploy
 (fp16 overflows the bf16-trained flan-t5-xl; int8 is size-only).
+
+## Training checkpoint
+
+`checkpoints/spamo/spamo_pjm_ms_dual_bleu5.36.ckpt` is a PyTorch Lightning checkpoint of SpaMo (PHOENIX-2014T init,
+PJM MS; epoch 43, dev BLEU-4 5.36). Its test run gave the 5.54 listed for `spamo-pjm-phoenix-ms`; which checkpoint
+the ONNX export was made from is not recorded.
+It is the starting point of the SpaMo-PJM `from_upstream` configs (v26 to v29). Moved here from the former
+`spamo-pjm-checkpoints` repo.