Commit History
Today
-
link the new paperweight home and default train.py to it
Verified signature6f8d8d2 -
remove the 3b-token experiment folder, now the same as v2
Verified signature3f7cc27
Yesterday
-
make the 3b-token run v2 and add the size comparison
Verified signaturea739c7e -
tinynext-v2-3b: 12.5M params, 56 min on GPU
Verified signaturedc714b2 -
export the compressed browser format from train.py instead of int8 activations
Verified signature4dece7f -
tinynext-v2-nofim: 12.5M params, 28 min on GPU
Verified signaturee5dde63 -
fetch the shared eval from v2/ even when uploading to an experiment folder
Verified signature0de7556 -
v2: 13 MB browser build (int8 download, float compute), export/compress scripts, onnx evals
Verified signature33bfb6d -
no-FIM models are evaluated with prefix-only prompts
Verified signature793c5c9 -
tinynext-v2: 12.5M params, 28 min on GPU
Verified signature1cffd47 -
prepare data on a CPU job, train from the prepared files
Verified signature63b64fd -
parallel data prep, data on GPU, trackio, safetensors
Verified signature9bba47b -
fix step count
Verified signature88610cb -
v2: training script, eval harness, eval set and baseline metrics
Verified signature3f86878