diff options
author | Tekky <98614666+xtekky@users.noreply.github.com> | 2024-10-03 13:20:46 +0200 |
---|---|---|
committer | GitHub <noreply@github.com> | 2024-10-03 13:20:46 +0200 |
commit | 6d19ba695655c85916deb2b6ba67c831fcf4c885 (patch) | |
tree | 467f56307f7a538b48eb9ffd6e318a9ff33e0159 /g4f/models.py | |
parent | ~ (diff) | |
parent | feat(g4f/models.py): enhance llama_3_1_405b with Blackbox provider (diff) | |
download | gpt4free-0.3.2.9.tar gpt4free-0.3.2.9.tar.gz gpt4free-0.3.2.9.tar.bz2 gpt4free-0.3.2.9.tar.lz gpt4free-0.3.2.9.tar.xz gpt4free-0.3.2.9.tar.zst gpt4free-0.3.2.9.zip |
Diffstat (limited to 'g4f/models.py')
-rw-r--r-- | g4f/models.py | 180 |
1 files changed, 120 insertions, 60 deletions
diff --git a/g4f/models.py b/g4f/models.py index 2940b96a..f1d50922 100644 --- a/g4f/models.py +++ b/g4f/models.py @@ -9,10 +9,8 @@ from .Provider import ( Allyfy, Bing, Binjie, - Bixin123, Blackbox, ChatGpt, - Chatgpt4o, Chatgpt4Online, ChatGptEs, ChatgptFree, @@ -80,12 +78,16 @@ default = Model( ReplicateHome, Upstage, Blackbox, - Bixin123, Binjie, Free2GPT, MagickPen, DeepInfraChat, LiteIcoding, + Airforce, + ChatHub, + Nexra, + ChatGptEs, + ChatHub, ]) ) @@ -106,7 +108,7 @@ gpt_35_turbo = Model( name = 'gpt-3.5-turbo', base_provider = 'OpenAI', best_provider = IterListProvider([ - Allyfy, Nexra, Bixin123, Airforce, + Allyfy, Nexra, Airforce, Liaobots, ]) ) @@ -115,7 +117,7 @@ gpt_4o = Model( name = 'gpt-4o', base_provider = 'OpenAI', best_provider = IterListProvider([ - Liaobots, Nexra, Airforce, Chatgpt4o, ChatGptEs, + Liaobots, Nexra, ChatGptEs, Airforce, OpenaiChat ]) ) @@ -133,7 +135,7 @@ gpt_4_turbo = Model( name = 'gpt-4-turbo', base_provider = 'OpenAI', best_provider = IterListProvider([ - Nexra, Bixin123, Liaobots, Airforce, Bing + Nexra, Liaobots, Airforce, Bing ]) ) @@ -141,9 +143,8 @@ gpt_4 = Model( name = 'gpt-4', base_provider = 'OpenAI', best_provider = IterListProvider([ - Nexra, Binjie, Airforce, - gpt_4_turbo.best_provider, gpt_4o.best_provider, gpt_4o_mini.best_provider, - Chatgpt4Online, Bing, OpenaiChat, + Nexra, Binjie, Airforce, Chatgpt4Online, Bing, OpenaiChat, + gpt_4_turbo.best_provider, gpt_4o.best_provider, gpt_4o_mini.best_provider ]) ) @@ -205,7 +206,7 @@ llama_3_1_70b = Model( llama_3_1_405b = Model( name = "llama-3.1-405b", base_provider = "Meta Llama", - best_provider = IterListProvider([Blackbox, DeepInfraChat, Airforce]) + best_provider = IterListProvider([DeepInfraChat, Blackbox, Airforce]) ) llama_3_1 = Model( @@ -214,12 +215,38 @@ llama_3_1 = Model( best_provider = IterListProvider([Nexra, llama_3_1_8b.best_provider, llama_3_1_70b.best_provider, llama_3_1_405b.best_provider,]) ) +# llama 3.2 +llama_3_2_11b = Model( + name = "llama-3.2-11b", + base_provider = "Meta Llama", + best_provider = IterListProvider([HuggingChat, HuggingFace]) +) + +llama_3_2_90b = Model( + name = "llama-3.2-90b", + base_provider = "Meta Llama", + best_provider = IterListProvider([Airforce]) +) + +# llamaguard +llamaguard_7b = Model( + name = "llamaguard-7b", + base_provider = "Meta Llama", + best_provider = IterListProvider([Airforce]) +) + +llamaguard_2_8b = Model( + name = "llamaguard-2-8b", + base_provider = "Meta Llama", + best_provider = IterListProvider([Airforce]) +) + ### Mistral ### mistral_7b = Model( name = "mistral-7b", base_provider = "Mistral", - best_provider = IterListProvider([HuggingChat, DeepInfraChat, Airforce, HuggingFace, DeepInfra]) + best_provider = IterListProvider([DeepInfraChat, Airforce, HuggingFace, DeepInfra]) ) mixtral_8x7b = Model( @@ -245,7 +272,13 @@ mistral_nemo = Model( mixtral_8x7b_dpo = Model( name = "mixtral-8x7b-dpo", base_provider = "NousResearch", - best_provider = Airforce + best_provider = IterListProvider([Airforce]) +) + +yi_34b = Model( + name = "yi-34b", + base_provider = "NousResearch", + best_provider = IterListProvider([Airforce]) ) hermes_3 = Model( @@ -310,49 +343,56 @@ gemma_2b = Model( ]) ) +# gemma 2 +gemma_2_27b = Model( + name = 'gemma-2-27b', + base_provider = 'Google', + best_provider = Airforce +) + gemma_2 = Model( name = 'gemma-2', base_provider = 'Google', - best_provider = ChatHub + best_provider = IterListProvider([ + ChatHub, + gemma_2_27b.best_provider, + ]) ) ### Anthropic ### -claude_2 = Model( - name = 'claude-2', - base_provider = 'Anthropic', - best_provider = You -) - -claude_2_0 = Model( - name = 'claude-2.0', +claude_2_1 = Model( + name = 'claude-2.1', base_provider = 'Anthropic', best_provider = Liaobots ) -claude_2_1 = Model( - name = 'claude-2.1', +claude_2 = Model( + name = 'claude-2', base_provider = 'Anthropic', - best_provider = Liaobots + best_provider = IterListProvider([ + You, + claude_2_1.best_provider, + ]) ) # claude 3 claude_3_opus = Model( name = 'claude-3-opus', base_provider = 'Anthropic', - best_provider = Liaobots + best_provider = IterListProvider([Airforce, Liaobots]) ) claude_3_sonnet = Model( name = 'claude-3-sonnet', base_provider = 'Anthropic', - best_provider = Liaobots + best_provider = IterListProvider([Airforce, Liaobots]) ) claude_3_haiku = Model( name = 'claude-3-haiku', base_provider = 'Anthropic', - best_provider = IterListProvider([DDG, Liaobots]) + best_provider = IterListProvider([DDG, Airforce, Liaobots]) ) claude_3 = Model( @@ -367,7 +407,7 @@ claude_3 = Model( claude_3_5_sonnet = Model( name = 'claude-3.5-sonnet', base_provider = 'Anthropic', - best_provider = IterListProvider([Blackbox, Liaobots]) + best_provider = IterListProvider([Blackbox, Airforce, Liaobots]) ) claude_3_5 = Model( @@ -417,15 +457,22 @@ command_r_plus = Model( sparkdesk_v1_1 = Model( name = 'sparkdesk-v1.1', base_provider = 'iFlytek', - best_provider = IterListProvider([FreeChatgpt, Airforce]) + best_provider = IterListProvider([FreeChatgpt]) ) ### Qwen ### +# qwen 1 +qwen_1_5_7b = Model( + name = 'qwen-1.5-7b', + base_provider = 'Qwen', + best_provider = Airforce +) + qwen_1_5_14b = Model( name = 'qwen-1.5-14b', base_provider = 'Qwen', - best_provider = FreeChatgpt + best_provider = IterListProvider([FreeChatgpt, Airforce]) ) qwen_1_5_72b = Model( @@ -440,22 +487,17 @@ qwen_1_5_110b = Model( best_provider = Airforce ) +# qwen 2 qwen_2_72b = Model( name = 'qwen-2-72b', base_provider = 'Qwen', best_provider = IterListProvider([DeepInfraChat, HuggingChat, Airforce, HuggingFace]) ) -qwen_turbo = Model( - name = 'qwen-turbo', - base_provider = 'Qwen', - best_provider = Bixin123 -) - qwen = Model( name = 'qwen', base_provider = 'Qwen', - best_provider = IterListProvider([Nexra, qwen_1_5_14b.best_provider, qwen_1_5_72b.best_provider, qwen_1_5_110b.best_provider, qwen_2_72b.best_provider, qwen_turbo.best_provider]) + best_provider = IterListProvider([Nexra, qwen_1_5_14b.best_provider, qwen_1_5_72b.best_provider, qwen_1_5_110b.best_provider, qwen_2_72b.best_provider]) ) @@ -488,13 +530,6 @@ yi_1_5_9b = Model( best_provider = FreeChatgpt ) -yi_34b = Model( - name = 'yi-34b', - base_provider = '01-ai', - best_provider = Airforce -) - - ### Upstage ### solar_1_mini = Model( name = 'solar-1-mini', @@ -542,14 +577,6 @@ wizardlm_2_8x22b = Model( best_provider = IterListProvider([DeepInfraChat, Airforce]) ) -### Together ### -sh_n_7b = Model( - name = 'sh-n-7b', - base_provider = 'Together', - best_provider = Airforce -) - - ### Yorickvp ### llava_13b = Model( name = 'llava-13b', @@ -611,7 +638,8 @@ grok_2_mini = Model( best_provider = Liaobots ) -# Perplexity AI + +### Perplexity AI ### sonar_online = Model( name = 'sonar-online', base_provider = 'Perplexity AI', @@ -625,6 +653,22 @@ sonar_chat = Model( ) +### Gryphe ### +mythomax_l2_13b = Model( + name = 'mythomax-l2-13b', + base_provider = 'Gryphe', + best_provider = IterListProvider([Airforce]) +) + + +### Pawan ### +cosmosrp = Model( + name = 'cosmosrp', + base_provider = 'Pawan', + best_provider = IterListProvider([Airforce]) +) + + ############# ### Image ### @@ -654,6 +698,7 @@ playground_v2_5 = Model( ) + ### Flux AI ### flux = Model( name = 'flux', @@ -805,6 +850,14 @@ class ModelUtils: 'llama-3.1-8b': llama_3_1_8b, 'llama-3.1-70b': llama_3_1_70b, 'llama-3.1-405b': llama_3_1_405b, + +# llama-3.2 +'llama-3.2-11b': llama_3_2_11b, +'llama-3.2-90b': llama_3_2_90b, + +# llamaguard +'llamaguard-7b': llamaguard_7b, +'llamaguard-2-8b': llamaguard_2_8b, ### Mistral ### @@ -835,12 +888,14 @@ class ModelUtils: 'gemma-2b': gemma_2b, 'gemma-2b-9b': gemma_2b_9b, 'gemma-2b-27b': gemma_2b_27b, + +# gemma-2 'gemma-2': gemma_2, +'gemma-2-27b': gemma_2_27b, ### Anthropic ### 'claude-2': claude_2, -'claude-2.0': claude_2_0, 'claude-2.1': claude_2_1, # claude 3 @@ -880,11 +935,11 @@ class ModelUtils: ### Qwen ### 'qwen': qwen, +'qwen-1.5-7b': qwen_1_5_7b, 'qwen-1.5-14b': qwen_1_5_14b, 'qwen-1.5-72b': qwen_1_5_72b, 'qwen-1.5-110b': qwen_1_5_110b, 'qwen-2-72b': qwen_2_72b, -'qwen-turbo': qwen_turbo, ### Zhipu AI ### @@ -908,11 +963,7 @@ class ModelUtils: ### DeepSeek ### 'deepseek': deepseek, - - -### Together ### -'sh-n-7b': sh_n_7b, - + ### Yorickvp ### 'llava-13b': llava_13b, @@ -947,9 +998,18 @@ class ModelUtils: 'grok-2': grok_2, 'grok-2-mini': grok_2_mini, + ### Perplexity AI ### 'sonar-online': sonar_online, 'sonar-chat': sonar_chat, + + +### Gryphe ### +'mythomax-l2-13b': sonar_chat, + + +### Pawan ### +'cosmosrp': cosmosrp, |