Views
No views yet
<think>...</think> reasoning traces in both
Romanian and English, alongside direct-answer mode.| Benchmark | Surogate-3.5-4B |
|---|---|
| Romanian text quality | |
| Invented word forms / 1k ↓ | 1.023 |
| English leakage / 1k ↓ | 0.783 |
| Missing diacritics / 1k ↓ | 0.342 |
| Knowledge & STEM | |
| MMLU-Pro | ~77.5 |
| MMLU-Redux | ~87.0 |
| GPQA Diamond | ~74.7 |
| SuperGPQA | ~51.8 |
| Instruction following | |
| RO IFEval prompt / instruction strict | 38.15 / 64.25 |
| EN IFEval prompt / instruction strict | 73.88 / 81.69 |
| IFEval | ~80.8 |
| IFBench | ~53.3 |
| MultiChallenge | ~44.1 |
| Math & reasoning | |
| RO GSM8K direct / thinking strict | 76.86 / 77.09 |
| EN GSM8K strict | 92.04 |
| HMMT Feb 25 / Nov 25 | ~76.2 / ~79.1 |
| PolyMATH | ~52.6 |
| Coding & agents | |
| LiveCodeBench v6 | ~54.7 |
| BFCL-V4 | ~49.3 |
| TAU2-Bench | ~78.3 |
| Long context | |
| LongBench v2 | ~49.0 |
| Multilingual & translation | |
| Translation EN to RO / RO to EN (chrF2) | 53.61 / 63.14 |
| WMT24++ | ~65.3 |
| MMMLU | ~74.6 |
| MMLU-ProX | ~70.1 |
| INCLUDE | ~69.6 |
| Global PIQA | ~77.3 |
~ are estimates derived
from the source foundation model's published results and measured capability
retention.enable_thinking=True, the reasoning trace follows the prompt language.Un tren parcurge 180 km în 3 ore. Care este viteza sa medie?1<think>
2Viteza medie este distanța împărțită la timp: 180 km / 3 h = 60 km/h.
3</think>
4Viteza medie este 60 km/h.A train travels 180 km in 3 hours. What is its average speed?1<think>
2Average speed is distance divided by time: 180 km / 3 h = 60 km/h.
3</think>
4The average speed is 60 km/h.1from transformers import AutoModelForImageTextToText, AutoProcessor
2
3model_id = "surogate/Surogate-3.5-4B"
4processor = AutoProcessor.from_pretrained(model_id)
5model = AutoModelForImageTextToText.from_pretrained(
6 model_id,
7 device_map="auto",
8 torch_dtype="auto",
9)
10
11messages = [
12 {"role": "user", "content": "Explică pe scurt de ce cerul este albastru."}
13]
14inputs = processor.apply_chat_template(
15 messages,
16 tokenize=True,
17 add_generation_prompt=True,
18 enable_thinking=False,
19 return_tensors="pt",
20 return_dict=True,
21).to(model.device)
22
23output = model.generate(**inputs, max_new_tokens=512)
24answer = processor.decode(
25 output[0][inputs.input_ids.shape[1]:],
26 skip_special_tokens=False,
27)
28print(answer)enable_thinking=True to request an explicit reasoning trace. The shipped
generation defaults are temperature 0.6, top-p 0.95, and top-k 20.