Views
No views yet
whiff-20M — это небольшая экспериментальная языковая модель на архитектуре Mamba2 с 20.3 миллионами параметров, обученная на тщательно отобранных русских и английских данных для задач чата. Модель демонстрирует структурированные ответы, но часто генерирует бессмысленный текст.1Mamba2Config(
2 vocab_size=8192,
3 hidden_size=512,
4 state_size=64,
5 num_heads=12,
6 num_hidden_layers=9,
7 conv_kernel=4,
8 expand=1.5,
9 n_groups=2
10)1from transformers import AutoTokenizer, AutoModelForCausalLM
2
3tokenizer = AutoTokenizer.from_pretrained("limloop/whiff-mamba2-20M")
4model = AutoModelForCausalLM.from_pretrained("limloop/whiff-mamba2-20M")
5
6def chat(messages, temp=0.5):
7 inputs = tokenizer.apply_chat_template(messages, return_tensors="pt")
8
9 outputs = model.generate(
10 inputs,
11 max_length=512,
12 top_k=40,
13 top_p=0.9,
14 repetition_penalty=1.1,
15 num_return_sequences=1,
16 temperature=temp,
17 do_sample=True,
18 eos_token_id=1
19 )
20
21 return tokenizer.decode(outputs[0], skip_special_tokens=True)
22
23# Пример
24dialog = [
25 {"role": "system", "content": "Ты — мудрый эльф."},
26 {"role": "user", "content": "Объясни квантовую физику."}
27]
28
29response = chat(dialog, temp=0.4)
30print(response)limloop/characters_dialogsIlyaGusev/gpt_roleplay_realmtamohannes/llm-roleplayradce/communication_datasetdatabricks/databricks-dolly-15kch1eph/RuGeoBenchnyuuzyou/ruschatgpt-qa0x22almostEvil/ru-riddles-3770x22almostEvil/tatoeba-mt-qna-oaDen4ikAI/ru_sberquad_long_answersVikhrmodels/GrandMaster-PRO-MAXHuggingFaceH4/ultrachat_200kOpenAssistant/oasst1OpenAssistant/oasst2PJMixers/hieunguyenminh_roleplay-deduped-ShareGPTArketov/hieunguyenminh_roleplay-deduped-ShareGPT_rulimloop/logic_duolimloop/ru_en_linguistic_exchangelimloop/multi_engagement_roleplay_corpuswhiff-20M is a small experimental language model based on the Mamba2 architecture with 20.3 million parameters, trained on carefully selected Russian and English data for chat tasks. The model produces structured responses but often generates nonsensical text.1Mamba2Config(
2 vocab_size=8192,
3 hidden_size=512,
4 state_size=64,
5 num_heads=12,
6 num_hidden_layers=9,
7 conv_kernel=4,
8 expand=1.5,
9 n_groups=2
10)1from transformers import AutoTokenizer, AutoModelForCausalLM
2
3tokenizer = AutoTokenizer.from_pretrained("limloop/whiff-mamba2-20M")
4model = AutoModelForCausalLM.from_pretrained("limloop/whiff-mamba2-20M")
5
6def chat(messages, temp=0.5):
7 inputs = tokenizer.apply_chat_template(messages, return_tensors="pt")
8
9 outputs = model.generate(
10 inputs,
11 max_length=512,
12 top_k=40,
13 top_p=0.9,
14 repetition_penalty=1.1,
15 num_return_sequences=1,
16 temperature=temp,
17 do_sample=True,
18 eos_token_id=1
19 )
20
21 return tokenizer.decode(outputs[0], skip_special_tokens=True)
22
23# Example
24dialog = [
25 {"role": "system", "content": "You are a wise elf."},
26 {"role": "user", "content": "Explain quantum physics."}
27]
28
29response = chat(dialog, temp=0.4)
30print(response)limloop/characters_dialogsIlyaGusev/gpt_roleplay_realmtamohannes/llm-roleplayradce/communication_datasetdatabricks/databricks-dolly-15kch1eph/RuGeoBenchnyuuzyou/ruschatgpt-qa0x22almostEvil/ru-riddles-3770x22almostEvil/tatoeba-mt-qna-oaDen4ikAI/ru_sberquad_long_answersVikhrmodels/GrandMaster-PRO-MAXHuggingFaceH4/ultrachat_200kOpenAssistant/oasst1OpenAssistant/oasst2PJMixers/hieunguyenminh_roleplay-deduped-ShareGPTArketov/hieunguyenminh_roleplay-deduped-ShareGPT_rulimloop/logic_duolimloop/ru_en_linguistic_exchangelimloop/multi_engagement_roleplay_corpus