Hugging Face's logo

IndexTeam
/
Index-Translate-9B

Translation
Safetensors
qwen3_5
index
Model card Files
xet
 Community
5
qijiv2000 commited on
Commit
d6bef93
·
verified ·
1 Parent(s): f25caaa

docs: full English model card from Index-Translate technical report

Browse files
Files changed (1) hide show
  1. README.md +91 -6
README.md CHANGED
@@ -1,16 +1,101 @@
1
  ---
2
- license: other
3
  tags:
4
  - translation
5
  - index
6
  ---
7
 
8
- # Index-MT-9B
9
 
10
- 通用翻译 9B 版。
 
 
11
 
12
- 在线体验：https://index-translate.bilibili.com （B站 Index-Translate 系列线上服务同款权重）。
13
 
14
- Base: Qwen3.5 系列。推理：vLLM 直接 `vllm serve` 本目录即可（safetensors + tokenizer 齐全）。
 
 
15
 
16
- > Model card 草稿，待补充评测数据与详细训练信息。
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
  ---
2
+ license: apache-2.0
3
  tags:
4
  - translation
5
  - index
6
  ---
7
 
8
+ # Index-Translate-9B
9
 
10
+ [![Demo](https://img.shields.io/badge/Demo-Online-orange.svg)](https://index-translate.bilibili.com/)
11
+ [![Code](https://img.shields.io/badge/GitHub-Index--Translate-yellow.svg)](https://github.com/bilibili/Index-Translate)
12
+ [![License](https://img.shields.io/badge/License-Apache--2.0-lightgrey.svg)](https://github.com/bilibili/Index-Translate/blob/main/LICENSE)
13
 
14
+ **Index-Translate-9B** is the 9B-parameter multilingual text translation model of the **Index-Translate** family, built on the Qwen3.5 backbone. It supports **148 languages** with multilingual instruction following — terminology, formatting, style, structure, and output-length constraints — and achieves translation quality comparable to 100B-scale translation models and frontier LLMs on the family's evaluation suite. It powers the [online demo](https://index-translate.bilibili.com/).
15
 
16
+ > 📄 Technical report: *Index-Translate: A Multilingual Translation Model Family — Text, Speech, Controlled Dubbing, and Long-Document Translation* (Index LLM Team, September 2026)
17
+ >
18
+ > 💻 Code, quickstarts, browser extension & video-dubbing pipeline: [github.com/bilibili/Index-Translate](https://github.com/bilibili/Index-Translate)
19
 
20
+ ## Highlights
21
+
22
+ - **148 languages** with instruction following: general translation, constrained translation (terminology, structure, style, context), and meme/internet-culture translation in a single checkpoint.
23
+ - **Best instTrans instruction-following score (IFscore 0.8209) among all compared models**, while remaining competitive on general translation.
24
+ - **Training recipe**: multilingual mid-training (general + monolingual + parallel + pivot-organized data) → specialist SFT (general / instruction / meme experts) → RL with judged rewards (GRPO; XCOMET-XXL + adequacy gates for general translation, Rubric-as-Reward for instruction following) under RIVAL adversarial judge supervision → **parameter interpolation** (model soup, general : instruction : meme = 0.8 : 0.1 : 0.1) → targeted **multi-teacher on-policy distillation (MOPD)**.
25
+ - **262,144-token context window**, greedy decoding recommended.
26
+ - Smaller sibling: [IndexTeam/Index-Translate-2B](https://huggingface.co/IndexTeam/Index-Translate-2B).
27
+
28
+ ## Benchmarks
29
+
30
+ Text translation (Table 4 of the report; WMT26 judge on a 0–100 scale, others 0–1):
31
+
32
+ | Model | FLORES COMET-22 | WMT24++ COMET-22 | WMT26 Judge | instTrans IFscore | IFMTBench IFscore | Vertical mean | MEME |
33
+ |---|---|---|---|---|---|---|---|
34
+ | **Index-Translate-9B** | **0.8789** | 0.8601 | 75.35 | **0.8209** | 0.8760 | 0.8451 | **0.7387** |
35
+ | Index-Translate-2B | 0.8655 | 0.8489 | 60.26 | 0.7569 | 0.7584 | 0.8377 | 0.6443 |
36
+ | Hy-MT2-7B | 0.8747 | 0.8593 | 60.51 | 0.6079 | 0.8741 | 0.8335 | 0.5139 |
37
+ | Hy-MT2-30B-A3B | 0.8787 | **0.8624** | 66.81 | 0.6415 | 0.9029 | **0.8459** | 0.5812 |
38
+ | North-Small-Translate (218B-A25B) | 0.8784 | 0.8578 | 68.37 | 0.5294 | 0.8635 | 0.8357 | 0.6836 |
39
+ | DeepSeek-V4.1-Flash | 0.8762 | 0.8510 | 83.55 | 0.6374 | 0.9090 | 0.8432 | 0.7424 |
40
+ | GPT-5.6-Sol | 0.8650 | 0.8469 | **89.10** | 0.7624 | **0.9367** | 0.8311 | 0.7194 |
41
+ | Qwen3.5-9B (base) | 0.8316 | 0.8073 | 60.31 | 0.0609 | 0.5980 | 0.8199 | 0.5728 |
42
+
43
+ Vertical-domain translation (COMET-22; GuoFeng uses document-COMET) — Index-Translate-9B leads the compared models on MuST-Cinema:
44
+
45
+ | Model | OPUS-Books | Biomedical | MTNT | MuST-Cinema | GuoFeng |
46
+ |---|---|---|---|---|---|
47
+ | **Index-Translate-9B** | 0.8010 | 0.8674 | 0.8246 | **0.8979** | **0.8345** |
48
+ | Index-Translate-2B | 0.7845 | 0.8635 | 0.8132 | 0.8930 | **0.8345** |
49
+
50
+ General capabilities (answer accuracy %) — exceeds Hy-MT2-7B on all four benchmarks:
51
+
52
+ | Model | C-Eval | GPQA-Diamond | INCLUDE | MMMLU |
53
+ |---|---|---|---|---|
54
+ | **Index-Translate-9B** | 69.6 | 36.3 | 63.1 | 66.6 |
55
+ | Hy-MT2-7B | 57.4 | 30.5 | 50.3 | 52.3 |
56
+ | Qwen3.5-9B (base) | 86.6 | 75.8 | 70.2 | 75.9 |
57
+
58
+ ## Usage
59
+
60
+ ```bash
61
+ vllm serve IndexTeam/Index-Translate-9B --served-model-name Index-Translate-9B --port 8000
62
+ ```
63
+
64
+ ```python
65
+ from openai import OpenAI
66
+
67
+ client = OpenAI(base_url="http://localhost:8000/v1", api_key="EMPTY")
68
+ resp = client.chat.completions.create(
69
+ model="Index-Translate-9B",
70
+ messages=[{"role": "user", "content": "Translate the following text into English:\n\n你好，世界！"}],
71
+ temperature=0,
72
+ )
73
+ print(resp.choices[0].message.content)
74
+ ```
75
+
76
+ Transformers and SGLang quickstarts: [inference/llm](https://github.com/bilibili/Index-Translate/tree/main/inference/llm).
77
+
78
+ ## Model family
79
+
80
+ | Model | Sizes | Task |
81
+ |---|---|---|
82
+ | **Index-Translate** | [2B](https://huggingface.co/IndexTeam/Index-Translate-2B) / [9B](https://huggingface.co/IndexTeam/Index-Translate-9B) | Multilingual text translation & translation instructions |
83
+ | Index-Echo S2TT | [2B](https://huggingface.co/IndexTeam/Index-Echo-S2TT-2B) / [9B](https://huggingface.co/IndexTeam/Index-Echo-S2TT-9B) | Speech-to-text translation |
84
+ | Index-Echo S2ST | [2B](https://huggingface.co/IndexTeam/Index-Echo-S2ST-2B) / [9B](https://huggingface.co/IndexTeam/Index-Echo-S2ST-9B) | Speech-to-speech translation with voice conditioning |
85
+ | Index-Homura | [2B](https://huggingface.co/IndexTeam/Index-Homura-2B) / [9B](https://huggingface.co/IndexTeam/Index-Homura-9B) | Syllable-controlled translation (dubbing) |
86
+ | Index-NaitiveLong (NaiLong) | [2B](https://huggingface.co/IndexTeam/Index-Nailong-2B) / [9B](https://huggingface.co/IndexTeam/Index-Nailong-9B) | Long-document translation |
87
+
88
+ ## Citation
89
+
90
+ ```bibtex
91
+ @techreport{indextranslate2026,
92
+ title={Index-Translate: A Multilingual Translation Model Family --- Text, Speech, Controlled Dubbing, and Long-Document Translation},
93
+ author={Index LLM Team},
94
+ year={2026},
95
+ month={September}
96
+ }
97
+ ```
98
+
99
+ ## License
100
+
101
+ Apache-2.0.