Update to latest training run; fix return

app.py

@@ -13,7 +13,7 @@ import os
 CLIP_PATH = "google/siglip-so400m-patch14-384"
 VLM_PROMPT = "A descriptive caption for this image:\n"
 MODEL_PATH = "meta-llama/Meta-Llama-3.1-8B"
-CHECKPOINT_PATH = Path("h2vtfhad")
+CHECKPOINT_PATH = Path("wpkklhc6")
 TITLE = "<h1><center>Foo</center></h1>"
 HF_TOKEN = os.environ.get("HF_TOKEN", None)
 
@@ -108,7 +108,7 @@ def stream_chat(input_image: Image.Image):
 
 	caption = tokenizer.batch_decode(generate_ids, skip_special_tokens=False, clean_up_tokenization_spaces=False)[0]
 
-	return [caption]
+	return caption.strip()
 
 
 with gr.Blocks() as demo:

{h2vtfhad → wpkklhc6}/config.yaml

@@ -3,7 +3,7 @@ device_batch_size: 2
 batch_size: 256
 learning_rate: 0.001
 warmup_samples: 18000
-max_samples: 500000
+max_samples: 600000
 save_every: 50000
 test_every: 50000
 use_amp: true
@@ -19,7 +19,7 @@ adam_beta2: 0.999
 adam_eps: 1.0e-08
 adam_weight_decay: 0.0
 clip_grad_norm: 1.0
-dataset: fancyfeast/joy-captioning-20240720a
+dataset: fancyfeast/joy-captioning-20240729a
 clip_model: google/siglip-so400m-patch14-384
 text_model: meta-llama/Meta-Llama-3.1-8B
 resume: null

{h2vtfhad → wpkklhc6}/image_adapter.pt

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:b06f70ede9f072d92855f91458a3ec09fdc27a3fddf859a4b1381cccff730fea
+oid sha256:2ebb1d1437bbb3264a6f25a896b25a7c7dd06c570c5de909dc2f19d3a5c5c110
 size 86018240