Commit History
-
Mark the repo as a quantization of the base model, not a finetune
Verified signatureHF was inferring base_model:finetune: from the base_model field alone. These are weight casts and a post-training quantization — no training was done — so base_model_relation: quantized is the accurate label.01747a7 -
Document the NVFP4 checkpoint with full precision comparisons
Verified signatureCovers the new 4-bit artifact: what NVFP4 is here (E2M1 in blocks of 16 with fp8 block scales, encoder Linears only, W4A16), directory contents and the pickle-free format, and how to load it. Adds the measurements: NVFP4 vs bf16 on 208 clips with paired bootstrap CIs and a per-slice breakdown, the memory table (0.78 GB peak load against bf16's 2.53 GB), the speed regression and the launch-bound reason for it, and the five other quantization schemes evaluated but not published — including NVFP4 W4A4, the one variant with a significant accuracy regression. States plainly that NVFP4 is not better than bf16: the corpus cannot separate them, and at 208 clips it cannot resolve differences below about 0.3 pp.1487e21 -
Add NVFP4 (4-bit) snapshot of parakeet-tdt-0.6b-v3
Verified signatureW4A16 NVFP4: the 217 encoder Linear layers quantized to 4-bit E2M1 in blocks of 16 with fp8 block scales; convolutions, LSTM decoder, joint network and all activations stay bf16. 0.51 GB of weights against bf16's 1.31 GB, and 0.78 GB peak VRAM against 1.43 GB. Over 208 clips WER is 2.86% against bf16's 2.99% — a difference the corpus cannot resolve (95% CI [-0.69, +0.30] pp, p=0.69). Decode is ~2.4x slower, so this is a memory trade, not a speed one. Loads with blurtd's PARAKEET_DTYPE=nvfp4. Packed weights in safetensors, quantization recipe in JSON — no pickles.8495b80 -
Add fp16 .nemo checkpoint (cast from upstream fp32)
Verified signaturee6585f4 -
Add bf16 .nemo checkpoint (cast from upstream fp32)
Verified signatureb00a8ef -
Add model card
Verified signaturee20feb8 -
initial commit
Verified signature327e775
