Views
No views yet
team-dentaku/dentaku-llm-jp-4-8b-stage1を公開しています。
システム全体の詳細はシステムレポートを参照してください。
| math-eval↑ | |
|---|---|
| team-dentaku/dentaku-llm-jp-4-8b-stage1 | 86.75 |
| team-dentaku/dentaku-llm-jp-4-8b-merged | 92.57 |
| team-dentaku/dentaku-llm-jp-4-8b-llama-pro-l40-stage1 | 89.96 |
| openai/gpt-oss-120b(教師モデル) | 92.57 |
trasformers==4.57.3で行っています．1import torch
2from transformers import AutoModelForCausalLM, AutoTokenizer
3
4def extract_code(text: str) -> str | None:
5 start_marker = "final<|message|>"
6 end_marker = "<|return|>"
7 start = text.find(start_marker)
8 if start == -1:
9 return None
10 start += len(start_marker)
11 end = text.find(end_marker, start)
12 if end == -1:
13 return None
14 return text[start:end].strip()
15
16
17
18model_name_or_path = "team-dentaku/dentaku-llm-jp-4-8b-stage1"
19question = r"$S_n$は次の式で与えられる。$S_n = \frac{1}{3^1} + \frac{2}{3^2} + \frac{3}{3^3} + \frac{4}{3^4} + \cdots + \frac{n}{3^n} = \sum_{k=1}^{n}\frac{k}{3^k}.$ このとき$\lim_{n \to \infty} S_n$の値を求めよ。"
20device = torch.device("cuda")
21
22tokenizer = AutoTokenizer.from_pretrained(model_name_or_path)
23model = AutoModelForCausalLM.from_pretrained(
24 model_name_or_path,
25 dtype=torch.bfloat16,
26 device_map=device,
27)
28model.eval()
29
30messages = [
31 [
32 {"role": "user", "content": question},
33 ]
34]
35
36input_ids = tokenizer.apply_chat_template(
37 messages,
38 return_tensors="pt",
39 add_generation_prompt=True,
40 tokenize=True
41).to(device)
42
43output_ids = model.generate(input_ids, max_new_tokens=1024, do_sample=False)
44output_texts = tokenizer.batch_decode(output_ids, skip_special_tokens=False)
45code = extract_code(output_texts[0])
46print(code)team-dentaku/dentaku-llm-jp-4-8b-stage1, which was trained using only the first-stage SFT dataset.
For details on the entire system, please refer to the system report.| math-eval↑ | |
|---|---|
| team-dentaku/dentaku-llm-jp-4-8b-stage1 | 86.75 |
| team-dentaku/dentaku-llm-jp-4-8b-merged | 92.57 |
| team-dentaku/dentaku-llm-jp-4-8b-llama-pro-l40-stage1 | 89.96 |
| openai/gpt-oss-120b (teacher model) | 92.57 |
transformers==4.57.3.1import torch
2from transformers import AutoModelForCausalLM, AutoTokenizer
3
4def extract_code(text: str) -> str | None:
5 start_marker = "final<|message|>"
6 end_marker = "<|return|>"
7 start = text.find(start_marker)
8 if start == -1:
9 return None
10 start += len(start_marker)
11 end = text.find(end_marker, start)
12 if end == -1:
13 return None
14 return text[start:end].strip()
15
16
17
18model_name_or_path = "team-dentaku/dentaku-llm-jp-4-8b-stage1"
19question = r"$S_n$は次の式で与えられる。$S_n = \frac{1}{3^1} + \frac{2}{3^2} + \frac{3}{3^3} + \frac{4}{3^4} + \cdots + \frac{n}{3^n} = \sum_{k=1}^{n}\frac{k}{3^k}.$ このとき$\lim_{n \to \infty} S_n$の値を求めよ。"
20device = torch.device("cuda")
21
22tokenizer = AutoTokenizer.from_pretrained(model_name_or_path)
23model = AutoModelForCausalLM.from_pretrained(
24 model_name_or_path,
25 dtype=torch.bfloat16,
26 device_map=device,
27)
28model.eval()
29
30messages = [
31 [
32 {"role": "user", "content": question},
33 ]
34]
35
36input_ids = tokenizer.apply_chat_template(
37 messages,
38 return_tensors="pt",
39 add_generation_prompt=True,
40 tokenize=True
41).to(device)
42
43output_ids = model.generate(input_ids, max_new_tokens=1024, do_sample=False)
44output_texts = tokenizer.batch_decode(output_ids, skip_special_tokens=False)
45code = extract_code(output_texts[0])
46print(code)