podcast-sponsor-remove, branch HEAD

try finetuning encoder-decoder model like t5gemma

2025-08-24T04:31:29Z

commit 0f5fcee95648d706b792affee0658e2e652fd904 parent 48cdc65c0eafa08fc39b5bfa49ef7c42943ac91d Author: vin Date: Sun, 24 Aug 2025 00:31:29 -0400 try finetuning encoder-decoder model like t5gemma Unfortunately once it finishes the finetuning process, it doesn't actually output anything. With decoder-only models like small gemma3, it also doesn't output anything meaningful. I'm tempted to just classify each token as an ad or not, finetune an encoder like some long-context BERT, and then a post-processing script to convert back to tokens and their timestamps. May end up being more performant and efficient too...

add finetuning scripts

2025-08-23T21:23:59Z

commit 48cdc65c0eafa08fc39b5bfa49ef7c42943ac91d parent 4ec10c33159f435963b73da5a9bd12f25f65cda9 Author: vin Date: Sat, 23 Aug 2025 17:23:59 -0400 add finetuning scripts

add helper script to split dataset.jsonl

2025-08-23T17:35:32Z

commit 4ec10c33159f435963b73da5a9bd12f25f65cda9 parent d5e69d1debc6df412f0862cb47ea90deebcd5dbf Author: vin Date: Sat, 23 Aug 2025 13:35:32 -0400 add helper script to split dataset.jsonl

add helper script for reviewing input into dataset

2025-08-23T17:29:32Z

commit d5e69d1debc6df412f0862cb47ea90deebcd5dbf parent d49d20dd810ed2dbcef6c926921b072713366654 Author: vin Date: Sat, 23 Aug 2025 13:29:32 -0400 add helper script for reviewing input into dataset

adapt asr.py for multiple file inputs

2025-08-23T06:26:35Z

commit d49d20dd810ed2dbcef6c926921b072713366654 parent 78decc2ddd05519780153a8d09f68884e8d11fc6 Author: vin Date: Sat, 23 Aug 2025 02:26:35 -0400 adapt asr.py for multiple file inputs

add highlight/visualizer script and update system prompt

2025-08-23T05:33:21Z

commit 78decc2ddd05519780153a8d09f68884e8d11fc6 parent 3b4865a270888643446d9092ba55d69b79117e5e Author: vin Date: Sat, 23 Aug 2025 01:33:21 -0400 add highlight/visualizer script and update system prompt

whisper.cpp replaced by onnx-asr

2025-08-23T03:28:51Z

commit 3b4865a270888643446d9092ba55d69b79117e5e parent a08bed429831b989b57b732ce04bf7158abc2cc1 Author: vin Date: Fri, 22 Aug 2025 23:28:51 -0400 whisper.cpp replaced by onnx-asr

update main script

2025-08-22T22:36:26Z

commit a08bed429831b989b57b732ce04bf7158abc2cc1 parent 87e22ba091dae8d81da690a405c15b453c61a0c5 Author: vin Date: Fri, 22 Aug 2025 18:36:26 -0400 update main script

add formatting

2025-08-22T21:51:28Z

commit 87e22ba091dae8d81da690a405c15b453c61a0c5 parent cc2d05f027ef2672a96d1b780e5af44b03375af8 Author: vin Date: Fri, 22 Aug 2025 17:51:28 -0400 add formatting

improve asr.py

2025-08-22T20:53:10Z

commit cc2d05f027ef2672a96d1b780e5af44b03375af8 parent 968414d2d38f3cb30ac4e1af3151a0842c113366 Author: vin Date: Fri, 22 Aug 2025 16:53:10 -0400 improve asr.py

update with newer parakeet

2025-08-22T19:47:24Z

commit 968414d2d38f3cb30ac4e1af3151a0842c113366 parent 8c01ad5cd727b4844e26e8a050c4094b545c6938 Author: vin Date: Fri, 22 Aug 2025 15:47:24 -0400 update with newer parakeet

add parakeet asr via onnx

2025-08-13T15:16:58Z

commit 8c01ad5cd727b4844e26e8a050c4094b545c6938 parent 94a64876146c9ca4707c5d750dfe1e464db41735 Author: vin Date: Wed, 13 Aug 2025 11:16:58 -0400 add parakeet asr via onnx found out that nvidia's parakeet has onnx support. it's about 4-5x faster than openai's whisper-large-v3-turbo (via whisper.cpp) and seems more accurate too without lots of errors near the end of the transcript. I thought I lost this repo, turns out I wasn't stupid and already had it pushed.

provide better examples in system prompt and retry on error

2025-07-19T21:59:23Z

commit 94a64876146c9ca4707c5d750dfe1e464db41735 parent 950c077309aff3036bdd2ab618983089faef41de Author: vin Date: Sat, 19 Jul 2025 17:59:23 -0400 provide better examples in system prompt and retry on error

save output in data/, not where podcasts are

2025-07-19T16:36:46Z

commit 950c077309aff3036bdd2ab618983089faef41de parent 2db3b781c15b45bc3f0771ad2ef30d6d88352adb Author: vin Date: Sat, 19 Jul 2025 12:36:46 -0400 save output in data/, not where podcasts are

change chunking to overlapping time segments

2025-07-19T15:41:37Z

commit 2db3b781c15b45bc3f0771ad2ef30d6d88352adb parent 534b7671e2d370bc498fa8d490121434b7a8e17f Author: vin Date: Sat, 19 Jul 2025 11:41:37 -0400 change chunking to overlapping time segments

output data in jsonl

2025-07-19T07:15:43Z

commit 534b7671e2d370bc498fa8d490121434b7a8e17f parent e4781586464f38f7b3ef1a663b2a50cca6cd2b56 Author: vin Date: Sat, 19 Jul 2025 03:15:43 -0400 output data in jsonl

respect ratelimit and change output extension to .gpt

2025-07-19T05:51:36Z

commit e4781586464f38f7b3ef1a663b2a50cca6cd2b56 parent bb8c6a7ed9b5b5c450c61408f432980be8e6b120 Author: vin Date: Sat, 19 Jul 2025 01:51:36 -0400 respect ratelimit and change output extension to .gpt

remove prototyping llm prints, use tqdm, use arguments as inputs

2025-07-19T05:16:12Z

commit bb8c6a7ed9b5b5c450c61408f432980be8e6b120 parent d9fdeaf5af49f065c2a5041d6e6230a2f371f53a Author: vin Date: Sat, 19 Jul 2025 01:16:12 -0400 remove prototyping llm prints, use tqdm, use arguments as inputs

initial payload

2025-07-19T02:45:53Z

commit d9fdeaf5af49f065c2a5041d6e6230a2f371f53a Author: vin Date: Fri, 18 Jul 2025 22:45:53 -0400 initial payload