gpt-oss-20b-balitanlp-cpt-bf16
A serving-ready copy of sapinsapin/gpt-oss-20b-balitanlp-cpt: gpt-oss-20b after continued pretraining on Filipino news (BalitaNLP).
The source repository keeps a float32 copy at its root (84 GB) and the BF16
copy in final_model/. This repository is that BF16 copy at the root, so
serving stacks load it directly:
| Source root | This repo | |
|---|---|---|
| Precision | float32 | BF16 (the training precision) |
| Size | 83.7 GB | 41.8 GB |
| Fits | multi-GPU | one 80 GB GPU (H100) with room for the KV cache |
Why not MXFP4
Stock gpt-oss-20b ships its expert weights in MXFP4 (4-bit), and it is
tempting to re-quantize a fine-tune the same way. For this checkpoint that
would undo the training. The continued pretraining moved the expert weights
by 0.12-0.33% - about one BF16 rounding step - and re-quantizing to MXFP4
snaps 99.8-100% of them back to the stock values (measured on layers 0-1
with check_finetune_survives.py in the NYO gateway repository,
scripts/custom-models). BF16 is the lowest precision that keeps what this
model learned.
Use
This is a base model after continued pretraining: it writes fluent Filipino but is not instruction-tuned.
vllm serve sapinsapin/gpt-oss-20b-balitanlp-cpt-bf16 \
--served-model-name gpt-oss-20b-balitanlp --max-model-len 32768
It is also available through NYO as
gpt-oss-20b-balitanlp when its endpoint is running.
Licence: Apache 2.0, as the source and base model.