Update README.md

README.md

@@ -15,6 +15,8 @@ license: cc-by-nc-4.0
 
 musicgen-songstarter-v0.2 is a [`musicgen-stereo-melody-large`](https://huggingface.co/facebook/musicgen-stereo-melody-large) fine-tuned on a dataset of melody loops from my Splice sample library. It's intended to be used to generate song ideas that are useful for music producers. It generates stereo audio in 32khz.
 
+**👀 Update:** I wrote a [blogpost](https://nateraw.com/posts/training_musicgen_songstarter.html) detailing how and why I trained this model, including training details, the dataset, Weights and Biases logs, etc.
+
 Compared to [`musicgen-songstarter-v0.1`](https://huggingface.co/nateraw/musicgen-songstarter-v0.1), this new version:
 - was trained on 3x more unique, manually-curated samples that I painstakingly purchased on Splice
 - Is twice the size, bumped up from size `medium` ➡️ `large` transformer LM
@@ -58,7 +60,7 @@ for idx, one_wav in enumerate(wav):
 Follow the following prompt format:
 
 ```
-{tag_1}, {tag_1}, ..., {tag_n}, {key}, {bpm} bpm
+{tag_1}, {tag_2}, ..., {tag_n}, {key}, {bpm} bpm
 ```
 
 For example:
@@ -67,6 +69,8 @@ For example:
 hip hop, soul, piano, chords, jazz, neo jazz, G# minor, 140 bpm
 ```
 
+For some example tags, [see the prompt format section of musicgen-songstarter-v0.1's readme](https://huggingface.co/nateraw/musicgen-songstarter-v0.1#prompt-format). The tags there are for the smaller v1 dataset, but should give you an idea of what the model saw.
+
 ## Samples
 
 <table style="width:100%; text-align:center;">
@@ -111,6 +115,23 @@ hip hop, soul, piano, chords, jazz, neo jazz, G# minor, 140 bpm
   </tr>
 </table>
 
+## Training Details
+
+For more verbose details, you can check out the [blogpost](https://nateraw.com/posts/training_musicgen_songstarter.html#training).
+
+- **code**:
+  - Repo is [here](https://github.com/nateraw/audiocraft). It's an undocumented fork of [facebookresearch/audiocraft](https://github.com/facebookresearch/audiocraft) where I rewrote the training loop with PyTorch Lightning, which worked a bit better for me.
+- **data**:
+  - around 1700-1800 samples I manually listened to + purchased via my personal [Splice](https://splice.com) account. About 7-8 hours of audio.
+  - Given the licensing terms, I cannot share the data.
+- **hardware**:
+  - 8xA100 40GB instance from Lambda Labs
+- **procedure**:
+  - trained for 10k steps, which took about 6 hours
+  - reduced segment duration at train time to 15 seconds
+- **hparams/logs**:
+  - See the wandb [run](https://wandb.ai/nateraw/musicgen-songstarter-v0.2/runs/63gh4l7m), which includes training metrics, logs, hardware metrics at train time, hyperparameters, and the exact command I used when I ran the training script.
+
 ## Acknowledgements
 
 This work would not have been possible without: