Add PROVENANCE.md - tensor-level provenance
Browse files- PROVENANCE.md +20 -0
PROVENANCE.md
ADDED
|
@@ -0,0 +1,20 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
# Tensor-level provenance
|
| 2 |
+
|
| 3 |
+
Merged checkpoint (NVFP4). Every weight in this model is NVIDIA's, copied from a
|
| 4 |
+
published NVIDIA checkpoint. No weight was trained, fine-tuned, distilled, or
|
| 5 |
+
recomputed here.
|
| 6 |
+
|
| 7 |
+
- `language_model.*` : nvidia/NVIDIA-Nemotron-3.5-Lightning-30B-A3B-NVFP4 (OpenMDW-1.1)
|
| 8 |
+
NVIDIA's published NVFP4 checkpoint, produced by PTQ and possibly QAD;
|
| 9 |
+
we did not determine which. No weight was recomputed by us.
|
| 10 |
+
- `vision_model.*`, `sound_encoder.*`, `mlp1.*`, `sound_projection.*` :
|
| 11 |
+
nvidia/Nemotron-3-Nano-Omni-30B-A3B-Reasoning-NVFP4 (NVIDIA Open Model Agreement)
|
| 12 |
+
BF16 in the source checkpoint; unquantized, copied byte-for-byte.
|
| 13 |
+
- modeling / configuration / tokenizer / processor code: Omni repo, unmodified.
|
| 14 |
+
|
| 15 |
+
Header-level rewrite only. No tensor was decoded, requantized or altered.
|
| 16 |
+
|
| 17 |
+
'Nemotron' is an NVIDIA trademark. OpenMDW-1.1 grants copyright, patent, database
|
| 18 |
+
and trade-secret rights and does NOT grant trademark rights; trademark use here is
|
| 19 |
+
limited to identifying the origin of the source models. See NOTICE for attribution
|
| 20 |
+
and LICENSE for the two governing agreements.
|