Upload VLA tokenizer with model card
Browse files- README.md +7 -5
- tokenizer.json +2 -2
README.md
CHANGED
|
@@ -14,9 +14,10 @@ license: apache-2.0
|
|
| 14 |
# VLA Tokenizer — Qwen3
|
| 15 |
|
| 16 |
Extended Qwen3 tokenizer for the **FineVideo-VLA** multimodal dataset.
|
| 17 |
-
Adds all VLA tokens (video, 3D pose, SNAC audio
|
|
|
|
| 18 |
|
| 19 |
-
**Vocab size: 257,
|
| 20 |
|
| 21 |
> **Why Qwen3?** The Qwen3 family has strong multilingual + reasoning abilities
|
| 22 |
> and native HuggingFace ecosystem support (vLLM, llama.cpp, transformers).
|
|
@@ -41,8 +42,9 @@ Adds all VLA tokens (video, 3D pose, SNAC audio) on top of the Qwen3 base tokeni
|
|
| 41 |
| SNAC Level 1 even | `<snac_132362>` – `<snac_136457>` | 4,096 | 25 Hz fine audio (even frames) |
|
| 42 |
| SNAC Level 1 odd | `<snac_144650>` – `<snac_148745>` | 4,096 | 25 Hz fine audio (odd frames) |
|
| 43 |
| SNAC wrappers | `<snac>`, `</snac>` | 2 | Block delimiters |
|
|
|
|
| 44 |
|
| 45 |
-
**Total VLA tokens added: 106,
|
| 46 |
|
| 47 |
---
|
| 48 |
|
|
@@ -60,7 +62,7 @@ Adds all VLA tokens (video, 3D pose, SNAC audio) on top of the Qwen3 base tokeni
|
|
| 60 |
from transformers import AutoTokenizer
|
| 61 |
|
| 62 |
tok = AutoTokenizer.from_pretrained("EmpathicRobotics/tokenizer-vla-qwen3")
|
| 63 |
-
print(len(tok)) #
|
| 64 |
|
| 65 |
# All VLA tokens are atomic
|
| 66 |
print(tok.encode("<seed2_1137>", add_special_tokens=False)) # single ID
|
|
@@ -113,7 +115,7 @@ tok.save_pretrained("tokenizer-vla-qwen3")
|
|
| 113 |
# vocab size: 257,897
|
| 114 |
```
|
| 115 |
|
| 116 |
-
Full script: `tools/build_tokenizers.py` in the
|
| 117 |
[finevideo-vla](https://github.com/TieuDaoChanNhan/finevideo-vla) repo.
|
| 118 |
|
| 119 |
---
|
|
|
|
| 14 |
# VLA Tokenizer — Qwen3
|
| 15 |
|
| 16 |
Extended Qwen3 tokenizer for the **FineVideo-VLA** multimodal dataset.
|
| 17 |
+
Adds all VLA tokens (video, 3D pose, SNAC audio, caption/speech wrappers) on
|
| 18 |
+
top of the Qwen3 base tokenizer.
|
| 19 |
|
| 20 |
+
**Vocab size: 257,901** (~151,669 Qwen3 base + 106,232 VLA tokens)
|
| 21 |
|
| 22 |
> **Why Qwen3?** The Qwen3 family has strong multilingual + reasoning abilities
|
| 23 |
> and native HuggingFace ecosystem support (vLLM, llama.cpp, transformers).
|
|
|
|
| 42 |
| SNAC Level 1 even | `<snac_132362>` – `<snac_136457>` | 4,096 | 25 Hz fine audio (even frames) |
|
| 43 |
| SNAC Level 1 odd | `<snac_144650>` – `<snac_148745>` | 4,096 | 25 Hz fine audio (odd frames) |
|
| 44 |
| SNAC wrappers | `<snac>`, `</snac>` | 2 | Block delimiters |
|
| 45 |
+
| Caption/speech wrappers | `<caption>`, `</caption>`, `<speech>`, `</speech>` | 4 | Inline caption/dialogue interleaving |
|
| 46 |
|
| 47 |
+
**Total VLA tokens added: 106,232**
|
| 48 |
|
| 49 |
---
|
| 50 |
|
|
|
|
| 62 |
from transformers import AutoTokenizer
|
| 63 |
|
| 64 |
tok = AutoTokenizer.from_pretrained("EmpathicRobotics/tokenizer-vla-qwen3")
|
| 65 |
+
print(len(tok)) # 257901
|
| 66 |
|
| 67 |
# All VLA tokens are atomic
|
| 68 |
print(tok.encode("<seed2_1137>", add_special_tokens=False)) # single ID
|
|
|
|
| 115 |
# vocab size: 257,897
|
| 116 |
```
|
| 117 |
|
| 118 |
+
Full script: `tools/tokenizer/build_tokenizers.py` in the
|
| 119 |
[finevideo-vla](https://github.com/TieuDaoChanNhan/finevideo-vla) repo.
|
| 120 |
|
| 121 |
---
|
tokenizer.json
CHANGED
|
@@ -1,3 +1,3 @@
|
|
| 1 |
version https://git-lfs.github.com/spec/v1
|
| 2 |
-
oid sha256:
|
| 3 |
-
size
|
|
|
|
| 1 |
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:d97a3c88b39e2dffeee8d470d767d48e2777c0d1395aaeee32b3709cb35e68a5
|
| 3 |
+
size 31655733
|