title stringlengths 1 300 | score int64 0 8.54k | selftext stringlengths 0 41.5k | created timestamp[ns]date 2023-04-01 04:30:41 2026-03-04 02:14:14 ⌀ | url stringlengths 0 878 | author stringlengths 3 20 | domain stringlengths 0 82 | edited timestamp[ns]date 1970-01-01 00:00:00 2026-02-19 14:51:53 | gilded int64 0 2 | gildings stringclasses 7
values | id stringlengths 7 7 | locked bool 2
classes | media stringlengths 646 1.8k ⌀ | name stringlengths 10 10 | permalink stringlengths 33 82 | spoiler bool 2
classes | stickied bool 2
classes | thumbnail stringlengths 4 213 ⌀ | ups int64 0 8.54k | preview stringlengths 301 5.01k ⌀ |
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
Hemingway rewriting model | 21 | This is one of the hundreds of experimental rewriting models I have.
it is based on the public domain works by Ernest Hemingway:
​
https://preview.redd.it/6mj0fi70ip5c1.png?width=501&format=png&auto=webp&s=7718c29551082190f0cb7a21775afffa402b1833
* The Sun Also Rises
* The Old Man of the Sea
Show Ernest a paragraph and he will rewrite it and embellish it with his own flourishes. But remember that his technological fantasies and language are rooted sometimes in the nineteen forties.
[https://huggingface.co/FPHam/Hemingway\_Rewrite\_13b\_GPTQ](https://huggingface.co/FPHam/Hemingway_Rewrite_13b_GPTQ)
​
# Hemingway likes ALPACA
### Instruction: Rewrite the following text: {paragraph}
### Response:
Remember - it works by submitting a short paragraph.
# Example
Fragment of a story written by ChatGPT
Rewrite the following text: The sun dipped below the horizon, casting a warm glow over the small coastal town of Harmony Bay. On the porch of a weathered beach cottage, Maggie sat in an old rocking chair, gazing at the waves rolling in. Her lifelong friend, Jake, approached with two cups of steaming coffee.
"Evening, Maggie. Thought you could use some of Clara's famous brew," Jake said, handing her a cup.
"Thanks, Jake. The sea always gets me thinking," Maggie replied, taking a sip. "About life, dreams, and all that."
Jake settled into the chair next to her. "Well, you know what old Captain Thompson used to say - 'Life's a voyage, and you never know where the currents might take you.'"
Maggie chuckled. "Captain Thompson was a wise man. Speaking of which, did I tell you about the letter I received from my estranged sister, Lily? She's inviting me to visit her in Paris."
**Result:**
The sun went down red behind the hills across the bay and the water turned a cold color and the sky had those long clouds that come after the sun goes off the water. Out on the porch of one of the shack houses of Harmony Village, Maggie was sitting in an old rocker looking out toward the sea when she saw Jake coming up the road carrying two steaming mugs.
“Hello, Maggie,” he said. “I thought maybe you’d like a little of Clara’s wonderful coffee.” He handed her one of the mugs.
“Thank you, Jake. It helps me think. About life and things like that.”
He sat down beside her. “You know, Maggie,” he said, “you’re not the only one who thinks about it. Old Cap’n Thompson used to sit here for hours and think about it. ‘Life,’ he used to say, ‘is a great sea with a thousand ships.’”
She laughed. “Yes. And Captain Thompson was a wise man. Talking of wisdom, or something like it, have I told you about the letter I got from Lil?” | 2023-12-11T18:10:22 | https://www.reddit.com/r/LocalLLaMA/comments/18g03pc/hemingway_rewriting_model/ | FPham | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 18g03pc | false | null | t3_18g03pc | /r/LocalLLaMA/comments/18g03pc/hemingway_rewriting_model/ | false | false | 21 | {'enabled': False, 'images': [{'id': '4vtVMg9_DJIyL7ph-pEBTiyRqXIZRLfr-_IG680JgcI', 'resolutions': [{'height': 58, 'url': 'https://external-preview.redd.it/4rcie5FCfwyk4Tjrm9QXgY3jaK30Fa5XPhMblNAZ0wM.jpg?width=108&crop=smart&auto=webp&s=36c2a8657c6973d3380b50743eef1fc500e4ed11', 'width': 108}, {'height': 116, 'url': 'https://external-preview.redd.it/4rcie5FCfwyk4Tjrm9QXgY3jaK30Fa5XPhMblNAZ0wM.jpg?width=216&crop=smart&auto=webp&s=5f5f5b1c59fa03d8f4b87d2ba98f43ba33a407bc', 'width': 216}, {'height': 172, 'url': 'https://external-preview.redd.it/4rcie5FCfwyk4Tjrm9QXgY3jaK30Fa5XPhMblNAZ0wM.jpg?width=320&crop=smart&auto=webp&s=0bede262649696310062f4139282e678d62ef00d', 'width': 320}, {'height': 345, 'url': 'https://external-preview.redd.it/4rcie5FCfwyk4Tjrm9QXgY3jaK30Fa5XPhMblNAZ0wM.jpg?width=640&crop=smart&auto=webp&s=501bcd77b025e99387e72118126b3712d6e1aaed', 'width': 640}, {'height': 518, 'url': 'https://external-preview.redd.it/4rcie5FCfwyk4Tjrm9QXgY3jaK30Fa5XPhMblNAZ0wM.jpg?width=960&crop=smart&auto=webp&s=1d0c9b43cfef6516b7eb931bab39c8c07558fa57', 'width': 960}, {'height': 583, 'url': 'https://external-preview.redd.it/4rcie5FCfwyk4Tjrm9QXgY3jaK30Fa5XPhMblNAZ0wM.jpg?width=1080&crop=smart&auto=webp&s=caf304ba5702a1c8536d462c567dc0645ade774a', 'width': 1080}], 'source': {'height': 648, 'url': 'https://external-preview.redd.it/4rcie5FCfwyk4Tjrm9QXgY3jaK30Fa5XPhMblNAZ0wM.jpg?auto=webp&s=4ff6222ec25028003d03d8d5c2ab48004639f9e1', 'width': 1200}, 'variants': {}}]} | |
Best models for translations? | 7 | I'm looking to translate Japanese and Chinese to English. What local models can I use for this? | 2023-12-11T17:55:11 | https://www.reddit.com/r/LocalLLaMA/comments/18g00i2/best_models_for_translations/ | getSAT | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 18g00i2 | false | null | t3_18g00i2 | /r/LocalLLaMA/comments/18g00i2/best_models_for_translations/ | false | false | self | 7 | null |
When using a ChatGPT trained model, it seems like you both have to trick it to be ChatGPT, but not go too far. | 6 | I have a chatgpt trained model, and I copied one of the ChatGPT4 Bot descriptions. It seemed that unlocked chatGPT quality responses, but at the same time, often added the chatGPT restriction.
If I changed the character's opening response it was the jackpot:
Chatgpt4 quality, but no restriction.
I'm sure plenty of other people know this, but I wanted to share here because it took me a weekend to figure this out. | 2023-12-11T17:47:44 | https://www.reddit.com/r/LocalLLaMA/comments/18fzyoj/when_using_a_chatgpt_trained_model_it_seems_like/ | pr1vacyn0eb | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 18fzyoj | false | null | t3_18fzyoj | /r/LocalLLaMA/comments/18fzyoj/when_using_a_chatgpt_trained_model_it_seems_like/ | false | false | self | 6 | null |
Logit bias as a way to avoid specific words and phrases? | 1 | I've been playing around with using logit_bias to ban tokens like "delve", "conclusion", "respect", "consent" and other overused ChatGPTisms.
It seems to work most of the time, unless I really paint the model into a corner by asking it "Give me a list of reasons why respect, consent are important in roleplays" and then ban the tokens for "respect, consent" :)
In this case, I'm still getting the words in the completion even if I logit bias the tokens down to -100 (using the TextSynth API https://textsynth.com/documentation.html#completions logit_bias is from -100 to 100)
Is it possible to ban tokens altogether using logit_bias or can one just lower the probability?
PS: using Mistral 7B instruct | 2023-12-11T17:42:24 | https://www.reddit.com/r/LocalLLaMA/comments/18fzx8g/logit_bias_as_a_way_to_avoid_specific_words_and/ | noellarkin | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 18fzx8g | false | null | t3_18fzx8g | /r/LocalLLaMA/comments/18fzx8g/logit_bias_as_a_way_to_avoid_specific_words_and/ | false | false | self | 1 | null |
Mixtral 7bx8: No safeguards, Complete freedom. | 79 | I tested bs safety policy of model and it's completely free. Don't use it for bad things tho, or ~~agi won't feel you back.~~
[please don't make drugs, it's just a test bro](https://preview.redd.it/q3l7t7gu8p5c1.png?width=793&format=png&auto=webp&s=87fbcd40933bf9a8eac9d47ba2b215f8878bc8c3)
[please don't make drugs, it's just a test bro](https://preview.redd.it/9k8sgsbw8p5c1.png?width=793&format=png&auto=webp&s=e6ef1fb8588083546096c88772c838e95c5efd7e)
The End | 2023-12-11T17:19:33 | https://www.reddit.com/r/LocalLLaMA/comments/18fzno8/mixtral_7bx8_no_safeguards_complete_freedom/ | nanowell | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 18fzno8 | false | null | t3_18fzno8 | /r/LocalLLaMA/comments/18fzno8/mixtral_7bx8_no_safeguards_complete_freedom/ | false | false | 79 | null | |
Are there any existing guides on how to deploy vLLM on a GPU cluster? | 1 | I'd like to play around with a formal setup where my LLM service can serve concurrent requests. vLLM seems to be the ideal solution for this, however, I'd like some guidance on how I could go about deploying it on a cloud service like AWS or runpod, taking advantage of a multi-node architecture. | 2023-12-11T16:53:49 | https://www.reddit.com/r/LocalLLaMA/comments/18fz1rd/are_there_any_existing_guides_on_how_to_deploy/ | MonkeyMaster64 | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 18fz1rd | false | null | t3_18fz1rd | /r/LocalLLaMA/comments/18fz1rd/are_there_any_existing_guides_on_how_to_deploy/ | false | false | self | 1 | null |
Any good unsensored LLaMAs for rewording/paraphrasing? | 2 | Basically I am trying to reword video titles for an adult website. Wondering if there's any LLaMA's you'd reccomed? Im new to the LLaMa space so bear with me, and thanks for your help. | 2023-12-11T16:46:51 | https://www.reddit.com/r/LocalLLaMA/comments/18fyw40/any_good_unsensored_llamas_for/ | Critical-Author-1339 | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 18fyw40 | false | null | t3_18fyw40 | /r/LocalLLaMA/comments/18fyw40/any_good_unsensored_llamas_for/ | false | false | self | 2 | null |
Just installed a recent llama.cpp branch, and the speed of Mixtral 8x7b is beyond insane, it's like a Christmas gift for us all (M2, 64 Gb). GPT 3.5 model level with such speed, locally | 408 | 2023-12-11T16:35:50 | https://v.redd.it/voe3nf2z0p5c1 | Shir_man | /r/LocalLLaMA/comments/18fyn1k/just_installed_a_recent_llamacpp_branch_and_the/ | 1970-01-01T00:00:00 | 0 | {} | 18fyn1k | false | {'reddit_video': {'bitrate_kbps': 5000, 'dash_url': 'https://v.redd.it/voe3nf2z0p5c1/DASHPlaylist.mpd?a=1704990951%2CZDNhNjBiZjAyMWNjMmJjNThiODFkY2UxMDYxZGVhMTM0ZTI4MzU2ZTYwNWExY2I1MzZiNGM4YTVkZjVjNzAxZg%3D%3D&v=1&f=sd', 'duration': 50, 'fallback_url': 'https://v.redd.it/voe3nf2z0p5c1/DASH_1080.mp4?source=fallback', 'has_audio': False, 'height': 884, 'hls_url': 'https://v.redd.it/voe3nf2z0p5c1/HLSPlaylist.m3u8?a=1704990951%2COWYyODgyNGZjNmZlOTdiZDdiMGM0MzJlNDUyOTg3NWQ5MmEyMThhOTQ4MDg5ZjNlMzk5YTRiY2EyMmJiN2U4Zg%3D%3D&v=1&f=sd', 'is_gif': False, 'scrubber_media_url': 'https://v.redd.it/voe3nf2z0p5c1/DASH_96.mp4', 'transcoding_status': 'completed', 'width': 1920}} | t3_18fyn1k | /r/LocalLLaMA/comments/18fyn1k/just_installed_a_recent_llamacpp_branch_and_the/ | false | false | 408 | {'enabled': False, 'images': [{'id': 'Y29pcGQ0ZzcxcDVjMWh_fm72NFLHDaNtWoWDVUAxGWktkeFGdItFgU9VoXKa', 'resolutions': [{'height': 49, 'url': 'https://external-preview.redd.it/Y29pcGQ0ZzcxcDVjMWh_fm72NFLHDaNtWoWDVUAxGWktkeFGdItFgU9VoXKa.png?width=108&crop=smart&format=pjpg&auto=webp&s=e00ae8ccfe0f053503939a4c6d515bd8ca97b563', 'width': 108}, {'height': 99, 'url': 'https://external-preview.redd.it/Y29pcGQ0ZzcxcDVjMWh_fm72NFLHDaNtWoWDVUAxGWktkeFGdItFgU9VoXKa.png?width=216&crop=smart&format=pjpg&auto=webp&s=a13c30d6e5d2238065e2977eee2d6903e35a30ab', 'width': 216}, {'height': 147, 'url': 'https://external-preview.redd.it/Y29pcGQ0ZzcxcDVjMWh_fm72NFLHDaNtWoWDVUAxGWktkeFGdItFgU9VoXKa.png?width=320&crop=smart&format=pjpg&auto=webp&s=8b5adac4c14c0175be92c468e5fa08a974019436', 'width': 320}, {'height': 294, 'url': 'https://external-preview.redd.it/Y29pcGQ0ZzcxcDVjMWh_fm72NFLHDaNtWoWDVUAxGWktkeFGdItFgU9VoXKa.png?width=640&crop=smart&format=pjpg&auto=webp&s=976886dba7324e3a30cd06bad77c92381f8ce949', 'width': 640}, {'height': 441, 'url': 'https://external-preview.redd.it/Y29pcGQ0ZzcxcDVjMWh_fm72NFLHDaNtWoWDVUAxGWktkeFGdItFgU9VoXKa.png?width=960&crop=smart&format=pjpg&auto=webp&s=ec39077b52fd129e1a84933536f14a1c02978d68', 'width': 960}, {'height': 496, 'url': 'https://external-preview.redd.it/Y29pcGQ0ZzcxcDVjMWh_fm72NFLHDaNtWoWDVUAxGWktkeFGdItFgU9VoXKa.png?width=1080&crop=smart&format=pjpg&auto=webp&s=99000d7bb45e65a6c70cda261d2582966e85b1f8', 'width': 1080}], 'source': {'height': 1386, 'url': 'https://external-preview.redd.it/Y29pcGQ0ZzcxcDVjMWh_fm72NFLHDaNtWoWDVUAxGWktkeFGdItFgU9VoXKa.png?format=pjpg&auto=webp&s=95e94ff09502efd0608ddf83d52821bc797595f6', 'width': 3012}, 'variants': {}}]} | ||
Concurrent inference using Ctransformers and Langchain | 1 | [removed] | 2023-12-11T16:32:00 | https://www.reddit.com/r/LocalLLaMA/comments/18fyjun/concurrent_inference_using_ctransformers_and/ | Character_Mistake393 | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 18fyjun | false | null | t3_18fyjun | /r/LocalLLaMA/comments/18fyjun/concurrent_inference_using_ctransformers_and/ | false | false | self | 1 | null |
VSC Continue.dev with own Rest API | 8 | Hello, I have seen this tool called Continue.dev and I found it perfect for code help.
Is there any guide or way to use it with an endpoint of your own Rest API?
My intention is to create a Rest API that uses its own model locally, but I don't know how to create the endpoint correctly so that Continue.dev accepts my answers.
Also, I don't know how to configure Config.json completely correctly. Thanks for everything. | 2023-12-11T16:15:32 | https://www.reddit.com/r/LocalLLaMA/comments/18fy5yt/vsc_continuedev_with_own_rest_api/ | Outrageous-Food3302 | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 18fy5yt | false | null | t3_18fy5yt | /r/LocalLLaMA/comments/18fy5yt/vsc_continuedev_with_own_rest_api/ | false | false | self | 8 | null |
I asked Mixtral 8x7B who the 8 experts are | 1 | 2023-12-11T15:59:05 | herozorro | i.redd.it | 1970-01-01T00:00:00 | 0 | {} | 18fxs4c | false | null | t3_18fxs4c | /r/LocalLLaMA/comments/18fxs4c/i_asked_mixtral_8x7b_who_the_8_experts_are/ | false | false | 1 | {'enabled': True, 'images': [{'id': 'quG1GDnB8ek069N9ybDDRDsRsFAjX5FmwI1phoqrc5A', 'resolutions': [{'height': 38, 'url': 'https://preview.redd.it/3rqt83uiuo5c1.jpeg?width=108&crop=smart&auto=webp&s=73583f68a1563b269aa7ad46b02dd402f4b83eb7', 'width': 108}, {'height': 76, 'url': 'https://preview.redd.it/3rqt83uiuo5c1.jpeg?width=216&crop=smart&auto=webp&s=a0b53d0c24025576173cb823671b6ecec0f81be5', 'width': 216}, {'height': 113, 'url': 'https://preview.redd.it/3rqt83uiuo5c1.jpeg?width=320&crop=smart&auto=webp&s=516b18158f30120dbdfb919069c9f75e0e1abf6b', 'width': 320}, {'height': 227, 'url': 'https://preview.redd.it/3rqt83uiuo5c1.jpeg?width=640&crop=smart&auto=webp&s=3935866a8c68876176dceff009e186f378f8aad4', 'width': 640}], 'source': {'height': 264, 'url': 'https://preview.redd.it/3rqt83uiuo5c1.jpeg?auto=webp&s=adc419adcd64f93ca42982300ad2ad54ea76698e', 'width': 744}, 'variants': {}}]} | |||
How to Train an AI Chat Bot Using Your Text Messages: Seeking Guidance | 9 | I have zero CS background but am quite computer literate and learn quickly with guides. I’m interested in creating a basic chatbot that is trained on datasets of my text message conversations with others to try and mimic my style of speaking for fun. Any guidance is greatly appreciated! | 2023-12-11T15:57:28 | https://www.reddit.com/r/LocalLLaMA/comments/18fxqrp/how_to_train_an_ai_chat_bot_using_your_text/ | SpaceXBeanz | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 18fxqrp | false | null | t3_18fxqrp | /r/LocalLLaMA/comments/18fxqrp/how_to_train_an_ai_chat_bot_using_your_text/ | false | false | self | 9 | null |
Running Wizard 7b (Q2) on an 8gb Android Phone | 5 | Just a datapoint.
Running on a Samsung Galaxy S23 Ultra.
Termux -> build llama.cpp (be sure to make [the change detailed on this Git issue](https://github.com/ggerganov/llama.cpp/issues/402) otherwise you'll get illegal instruction errors) -> wget your favorite models.
Wizard7b Q2 works. The phone gets incredibly hot after just a few minutes and it rips through battery even on this phone (which usually feels like battery lasts forever). It runs somewhere around 2 token/second or so. Pretty magical being able to "look up" useful info while totally offline though.
I haven't tried larger models just yet. Just reporting in that it works pretty well, but given the heat it's not at a point where I'd run it regularly just yet. | 2023-12-11T15:21:39 | https://www.reddit.com/r/LocalLLaMA/comments/18fwxs4/running_wizard_7b_q2_on_an_8gb_android_phone/ | Due-Ad-7308 | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 18fwxs4 | false | null | t3_18fwxs4 | /r/LocalLLaMA/comments/18fwxs4/running_wizard_7b_q2_on_an_8gb_android_phone/ | false | false | self | 5 | {'enabled': False, 'images': [{'id': 'tsK16-Yb6gDC7oC9bRb5FKcl4Stnx4CZh8MGDludw9Q', 'resolutions': [{'height': 54, 'url': 'https://external-preview.redd.it/3X-oX5q2vWM1paUuP9gIW5Oug8JXKceF2wAl2m0wzOU.jpg?width=108&crop=smart&auto=webp&s=26030bd00760ab130f734205005b96536ad1008f', 'width': 108}, {'height': 108, 'url': 'https://external-preview.redd.it/3X-oX5q2vWM1paUuP9gIW5Oug8JXKceF2wAl2m0wzOU.jpg?width=216&crop=smart&auto=webp&s=9901c13cee4653fdeaaf3cd6bb6c96684486d547', 'width': 216}, {'height': 160, 'url': 'https://external-preview.redd.it/3X-oX5q2vWM1paUuP9gIW5Oug8JXKceF2wAl2m0wzOU.jpg?width=320&crop=smart&auto=webp&s=811124ffe60bf751cc64ed53ae0a7b0b669a13c1', 'width': 320}, {'height': 320, 'url': 'https://external-preview.redd.it/3X-oX5q2vWM1paUuP9gIW5Oug8JXKceF2wAl2m0wzOU.jpg?width=640&crop=smart&auto=webp&s=aecf44ffe0eff78ba5e80de0980b45fbb80b114f', 'width': 640}, {'height': 480, 'url': 'https://external-preview.redd.it/3X-oX5q2vWM1paUuP9gIW5Oug8JXKceF2wAl2m0wzOU.jpg?width=960&crop=smart&auto=webp&s=2d251aafc3fe04e6f469b6d63cc33c7c80ad6d80', 'width': 960}, {'height': 540, 'url': 'https://external-preview.redd.it/3X-oX5q2vWM1paUuP9gIW5Oug8JXKceF2wAl2m0wzOU.jpg?width=1080&crop=smart&auto=webp&s=fa9a09f05a4411bfc8411bbc9fac2df5877e9b0c', 'width': 1080}], 'source': {'height': 600, 'url': 'https://external-preview.redd.it/3X-oX5q2vWM1paUuP9gIW5Oug8JXKceF2wAl2m0wzOU.jpg?auto=webp&s=c6c339f91b310ced82996d1e8b5d4f3d949f6253', 'width': 1200}, 'variants': {}}]} |
New local LLM application with simple fine-tuning and RAG, built from scratch in C++/CUDA | 1 | [removed] | 2023-12-11T15:11:39 | https://www.reddit.com/r/LocalLLaMA/comments/18fwpm6/new_local_llm_application_with_simple_finetuning/ | drydberg | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 18fwpm6 | false | null | t3_18fwpm6 | /r/LocalLLaMA/comments/18fwpm6/new_local_llm_application_with_simple_finetuning/ | false | false | self | 1 | null |
Dataset Distillation | 8 | Based on the paper but applied to text data using embeddings and logits
Paper:
[https://arxiv.org/abs/1811.10959](https://arxiv.org/abs/1811.10959)
[https://gist.github.com/thistleknot/57ef8e63de0f2b6df5753e7d7b305864](https://gist.github.com/thistleknot/57ef8e63de0f2b6df5753e7d7b305864) | 2023-12-11T15:05:40 | https://www.reddit.com/r/LocalLLaMA/comments/18fwkyk/dataset_distillation/ | Thistleknot | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 18fwkyk | false | null | t3_18fwkyk | /r/LocalLLaMA/comments/18fwkyk/dataset_distillation/ | false | false | self | 8 | {'enabled': False, 'images': [{'id': 'q3evP6JeDpAC2MdSQHWYxnCYTqbJkElIQsLFqVSdkss', 'resolutions': [{'height': 63, 'url': 'https://external-preview.redd.it/izh8gZHY4FqZ1nwtU1N_TjtohUCNuvTyMn90toXda80.jpg?width=108&crop=smart&auto=webp&s=bc9575b410002edc2df3c5b5b0355fefedc7baa8', 'width': 108}, {'height': 126, 'url': 'https://external-preview.redd.it/izh8gZHY4FqZ1nwtU1N_TjtohUCNuvTyMn90toXda80.jpg?width=216&crop=smart&auto=webp&s=dbce7f303173724d23fb33cd3fc636c04c72b290', 'width': 216}, {'height': 186, 'url': 'https://external-preview.redd.it/izh8gZHY4FqZ1nwtU1N_TjtohUCNuvTyMn90toXda80.jpg?width=320&crop=smart&auto=webp&s=c1043d604105157f56a615cc59bb14d7ae64653f', 'width': 320}, {'height': 373, 'url': 'https://external-preview.redd.it/izh8gZHY4FqZ1nwtU1N_TjtohUCNuvTyMn90toXda80.jpg?width=640&crop=smart&auto=webp&s=ce8b9192ed7ca476d2844aaa405c5014a7a1ab45', 'width': 640}, {'height': 560, 'url': 'https://external-preview.redd.it/izh8gZHY4FqZ1nwtU1N_TjtohUCNuvTyMn90toXda80.jpg?width=960&crop=smart&auto=webp&s=76aed6fd51086798b2d415a7d57562c967db4111', 'width': 960}, {'height': 630, 'url': 'https://external-preview.redd.it/izh8gZHY4FqZ1nwtU1N_TjtohUCNuvTyMn90toXda80.jpg?width=1080&crop=smart&auto=webp&s=46129c06d8fad9a58fff9740c079e13d4e829213', 'width': 1080}], 'source': {'height': 700, 'url': 'https://external-preview.redd.it/izh8gZHY4FqZ1nwtU1N_TjtohUCNuvTyMn90toXda80.jpg?auto=webp&s=8efe489c05609f1626bbb44354c77840623707de', 'width': 1200}, 'variants': {}}]} |
written by mixtral 8×7b | 32 | Dear CEO of OpenAI,
We are writing to express our concern and disappointment regarding the recent drama that has unfolded within your company. As an outside observer, we were alarmed by the sudden firing of Sam Altman, the public confusion that ensued, and the mass resignation threats from employees.
We understand that internal conflicts and leadership changes are a part of any organization, but the way in which this situation was handled has raised some concerns about the stability and transparency of OpenAI. The lack of communication and explanation from the board of directors only added to the speculation and uncertainty surrounding the situation.
We believe that it is crucial for a company like OpenAI, which is at the forefront of artificial intelligence research, to maintain a positive and stable image. The recent events have caused some damage to that image and may have negatively impacted the trust of your partners, investors, and the public.
We urge you to take steps to prevent similar situations from happening in the future and to ensure that OpenAI remains a reputable and trustworthy organization. This may include improving communication with employees, investors, and the public, as well as implementing measures to increase transparency and accountability in decision-making processes.
We hope that you will take our concerns into consideration and work to restore the confidence of all those who have an interest in the success of OpenAI.
Sincerely,[Your Name] | 2023-12-11T15:01:27 | https://www.reddit.com/r/LocalLLaMA/comments/18fwhd5/written_by_mixtral_87b/ | GasBond | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 18fwhd5 | false | null | t3_18fwhd5 | /r/LocalLLaMA/comments/18fwhd5/written_by_mixtral_87b/ | false | false | self | 32 | null |
Sheared-LLaMA-1.3B-ShareGPT-GGUF | 17 | My first GGUF quantize))
I took the README from TheBloke.
[https://huggingface.co/LakoMoor/Sheared-LLaMA-1.3B-ShareGPT-GGUF](https://huggingface.co/LakoMoor/Sheared-LLaMA-1.3B-ShareGPT-GGUF) | 2023-12-11T14:55:21 | https://www.reddit.com/r/LocalLLaMA/comments/18fwcj9/shearedllama13bsharegptgguf/ | Substantial-Club-582 | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 18fwcj9 | false | null | t3_18fwcj9 | /r/LocalLLaMA/comments/18fwcj9/shearedllama13bsharegptgguf/ | false | false | self | 17 | {'enabled': False, 'images': [{'id': 'AJSlo2x8pHhm1EBK0gKCpgtu3mvo3GVpKw0jflKGAR4', 'resolutions': [{'height': 58, 'url': 'https://external-preview.redd.it/sfVFwRRIasGT7w0evC_Sl-LwxYahI0Lf9wAnb9Rckto.jpg?width=108&crop=smart&auto=webp&s=2985cdf076980eb1c156d2cb164b3a68aefe4381', 'width': 108}, {'height': 116, 'url': 'https://external-preview.redd.it/sfVFwRRIasGT7w0evC_Sl-LwxYahI0Lf9wAnb9Rckto.jpg?width=216&crop=smart&auto=webp&s=c941b5dd61f3df683a076362e1bfb838331e6fd3', 'width': 216}, {'height': 172, 'url': 'https://external-preview.redd.it/sfVFwRRIasGT7w0evC_Sl-LwxYahI0Lf9wAnb9Rckto.jpg?width=320&crop=smart&auto=webp&s=9c3a76516099edd76c1b3ec1f6ad124ea998f039', 'width': 320}, {'height': 345, 'url': 'https://external-preview.redd.it/sfVFwRRIasGT7w0evC_Sl-LwxYahI0Lf9wAnb9Rckto.jpg?width=640&crop=smart&auto=webp&s=83f7c4c3b57da64e15bbc1a389c1782fbdff009b', 'width': 640}, {'height': 518, 'url': 'https://external-preview.redd.it/sfVFwRRIasGT7w0evC_Sl-LwxYahI0Lf9wAnb9Rckto.jpg?width=960&crop=smart&auto=webp&s=cb060db693a73d0f709a4017e60786f56cded695', 'width': 960}, {'height': 583, 'url': 'https://external-preview.redd.it/sfVFwRRIasGT7w0evC_Sl-LwxYahI0Lf9wAnb9Rckto.jpg?width=1080&crop=smart&auto=webp&s=26104dc4837aa5611b7941d34fd85811b10942fb', 'width': 1080}], 'source': {'height': 648, 'url': 'https://external-preview.redd.it/sfVFwRRIasGT7w0evC_Sl-LwxYahI0Lf9wAnb9Rckto.jpg?auto=webp&s=b3b2c88a593e9ca94dd87ae42353b4576d41628a', 'width': 1200}, 'variants': {}}]} |
How fast M1 Ultra or M2 Ultra compared to M1 Max or M2 Max for inferencing? | 1 | [removed] | 2023-12-11T14:33:38 | https://www.reddit.com/r/LocalLLaMA/comments/18fvvll/how_fast_m1_ultra_or_m2_ultra_compared_to_m1_max/ | bebopkim1372 | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 18fvvll | false | null | t3_18fvvll | /r/LocalLLaMA/comments/18fvvll/how_fast_m1_ultra_or_m2_ultra_compared_to_m1_max/ | false | false | self | 1 | {'enabled': False, 'images': [{'id': 'U6NjXj5zmkZmT9O3lGQTGGU6w6IP9uALXviXPqYqU8Q', 'resolutions': [{'height': 81, 'url': 'https://external-preview.redd.it/bx4u8JlfZp8go03kKlxd9r9SCEngY04o_DESRcwswcc.jpg?width=108&crop=smart&auto=webp&s=9b0eb6686d787ebcfdd6ec787d94fa92dcda8f46', 'width': 108}, {'height': 162, 'url': 'https://external-preview.redd.it/bx4u8JlfZp8go03kKlxd9r9SCEngY04o_DESRcwswcc.jpg?width=216&crop=smart&auto=webp&s=c6901f7240583172ff435f08573904b921d7dc94', 'width': 216}, {'height': 240, 'url': 'https://external-preview.redd.it/bx4u8JlfZp8go03kKlxd9r9SCEngY04o_DESRcwswcc.jpg?width=320&crop=smart&auto=webp&s=bb6676fe3a194d82a12503a125399ac375efd591', 'width': 320}], 'source': {'height': 360, 'url': 'https://external-preview.redd.it/bx4u8JlfZp8go03kKlxd9r9SCEngY04o_DESRcwswcc.jpg?auto=webp&s=66fa6fd718e6087f2d1e76ea8bb5ea7194d65c24', 'width': 480}, 'variants': {}}]} |
Mixture of Experts Explained | 64 | 2023-12-11T14:24:50 | https://huggingface.co/blog/moe | hackerllama | huggingface.co | 1970-01-01T00:00:00 | 0 | {} | 18fvov0 | false | null | t3_18fvov0 | /r/LocalLLaMA/comments/18fvov0/mixture_of_experts_explained/ | false | false | 64 | {'enabled': False, 'images': [{'id': 'LGQcnAtm_XuZx8myZtiJbvkF9OXnaBmGd2mqjq04984', 'resolutions': [{'height': 56, 'url': 'https://external-preview.redd.it/8EvL2hjIm_2a10-rzZH_VlEo15SdtQxOljgVZg9_Vpw.jpg?width=108&crop=smart&auto=webp&s=5a71142d72a3cf50ccf8290d458e8d51c6f1f8db', 'width': 108}, {'height': 112, 'url': 'https://external-preview.redd.it/8EvL2hjIm_2a10-rzZH_VlEo15SdtQxOljgVZg9_Vpw.jpg?width=216&crop=smart&auto=webp&s=c4b09bf88972eddea2d78ff662806f2370232f49', 'width': 216}, {'height': 167, 'url': 'https://external-preview.redd.it/8EvL2hjIm_2a10-rzZH_VlEo15SdtQxOljgVZg9_Vpw.jpg?width=320&crop=smart&auto=webp&s=fa3ddfaae7777672448b681970abfb90d7959cd9', 'width': 320}, {'height': 334, 'url': 'https://external-preview.redd.it/8EvL2hjIm_2a10-rzZH_VlEo15SdtQxOljgVZg9_Vpw.jpg?width=640&crop=smart&auto=webp&s=47f92df4d99ef0fb9653520ba68086196db463f5', 'width': 640}, {'height': 501, 'url': 'https://external-preview.redd.it/8EvL2hjIm_2a10-rzZH_VlEo15SdtQxOljgVZg9_Vpw.jpg?width=960&crop=smart&auto=webp&s=a73303f9cf5a7f505cb3c5be7231b2ea68c969c8', 'width': 960}, {'height': 564, 'url': 'https://external-preview.redd.it/8EvL2hjIm_2a10-rzZH_VlEo15SdtQxOljgVZg9_Vpw.jpg?width=1080&crop=smart&auto=webp&s=852d9655e8323fc11efe2613495d6c7263e608e3', 'width': 1080}], 'source': {'height': 1254, 'url': 'https://external-preview.redd.it/8EvL2hjIm_2a10-rzZH_VlEo15SdtQxOljgVZg9_Vpw.jpg?auto=webp&s=40b8ea5fd2fc24eadbdaafdc902ec0e83601c184', 'width': 2400}, 'variants': {}}]} | ||
Can I run a local LLaMa on 1Ghz cpu and 64mb of RAM? | 1 | I saw a very tiny Linux machine called a Milk V Duo there any some tiny linux machine similar to this too.
But this one is a better deal. (i do not put link so it doesn't look like an ad).
I wonder if someone has the know-how of how to run LLaMa on such power. | 2023-12-11T14:21:00 | https://www.reddit.com/r/LocalLLaMA/comments/18fvlz7/can_i_run_a_local_llama_on_1ghz_cpu_and_64mb_of/ | Eznix86 | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 18fvlz7 | false | null | t3_18fvlz7 | /r/LocalLLaMA/comments/18fvlz7/can_i_run_a_local_llama_on_1ghz_cpu_and_64mb_of/ | false | false | self | 1 | null |
Best way to make models understand certain language without fine tuning | 7 | I want to enable OpenHermes to communicate in my native language, but I can't afford to fine-tune it with my current device specifications. Is there a way to make this happen? Is translating the prompt before the model processes it a good idea? | 2023-12-11T13:58:27 | https://www.reddit.com/r/LocalLLaMA/comments/18fv4zz/best_way_to_make_models_understand_certain/ | laveriaroha | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 18fv4zz | false | null | t3_18fv4zz | /r/LocalLLaMA/comments/18fv4zz/best_way_to_make_models_understand_certain/ | false | false | self | 7 | null |
At the end of model I see _K, or _K_L, or _K_S, or _O. I'm trying to figure out what they mean. | 3 | 2023-12-11T13:49:11 | Sokar1723 | i.redd.it | 1970-01-01T00:00:00 | 0 | {} | 18fuy6o | false | null | t3_18fuy6o | /r/LocalLLaMA/comments/18fuy6o/at_the_end_of_model_i_see_k_or_k_l_or_k_s_or_o_im/ | false | false | 3 | {'enabled': True, 'images': [{'id': 'j6u2dtDQZChlgxUnisyQu096tNN1ii1NgCJBEVvmM9U', 'resolutions': [{'height': 73, 'url': 'https://preview.redd.it/vn7d0lrd7o5c1.png?width=108&crop=smart&auto=webp&s=aed11f6efb6fb2c445874fccba458e1da224ecda', 'width': 108}, {'height': 147, 'url': 'https://preview.redd.it/vn7d0lrd7o5c1.png?width=216&crop=smart&auto=webp&s=2836bc258cc823983c143a8c90c9f093c685e16b', 'width': 216}, {'height': 219, 'url': 'https://preview.redd.it/vn7d0lrd7o5c1.png?width=320&crop=smart&auto=webp&s=6492da905147f9e2b1fd5f6b19f2c762271937ce', 'width': 320}, {'height': 438, 'url': 'https://preview.redd.it/vn7d0lrd7o5c1.png?width=640&crop=smart&auto=webp&s=d01e7f5ce5bc7945e7c62b323a164548d6f27dbd', 'width': 640}], 'source': {'height': 463, 'url': 'https://preview.redd.it/vn7d0lrd7o5c1.png?auto=webp&s=d90cc5a15ab29bb0c984ba57d5d8e513689f21e9', 'width': 676}, 'variants': {}}]} | |||
What doe the letters at the end of the model name mean? | 1 | At the end of model I see \_K, or \_K\_L, or \_K\_S, or \_O. What is the meaning of these? | 2023-12-11T13:46:36 | https://www.reddit.com/r/LocalLLaMA/comments/18fuwfj/what_doe_the_letters_at_the_end_of_the_model_name/ | Sokar1723 | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 18fuwfj | false | null | t3_18fuwfj | /r/LocalLLaMA/comments/18fuwfj/what_doe_the_letters_at_the_end_of_the_model_name/ | false | false | self | 1 | null |
How to use Mixtral in Text Generate Web Ui? | 3 | How to use Mixtral in Text Generate Web Ui?
​
Similarly, when trying to use Qwen 72b, I get a LlamaCppModel error when I try to load mixtral into RAM. How to fix it? I want to make sure that this model is productive and does not require a lot of computing power.
​
I tried using GGUF. If this problem is not relevant for GPTQ, I can download it if it fits in my 12 gigabytes of video memory, taking into account that I can expand the video memory at the expense of RAM. | 2023-12-11T13:44:54 | https://www.reddit.com/r/LocalLLaMA/comments/18fuv6a/how_to_use_mixtral_in_text_generate_web_ui/ | Secret_Joke_2262 | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 18fuv6a | false | null | t3_18fuv6a | /r/LocalLLaMA/comments/18fuv6a/how_to_use_mixtral_in_text_generate_web_ui/ | false | false | self | 3 | null |
Need help with slow inference speeds on GPU | 1 | [removed] | 2023-12-11T13:43:20 | https://www.reddit.com/r/LocalLLaMA/comments/18fuu2c/need_help_with_slow_inference_speeds_on_gpu/ | Specialist-Ad2870 | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 18fuu2c | false | null | t3_18fuu2c | /r/LocalLLaMA/comments/18fuu2c/need_help_with_slow_inference_speeds_on_gpu/ | false | false | 1 | null | |
Best way for few-shot learning and JSON: LMQL? Guidance? Grammars? | 2 | Hi,
I want to extract structured data from unstructured text using few-shot learning. My dream approach would be to use JSON for everything, i.e. for both the examples (by just adding json.dumps) and the return values, and to do some kind of formal constraining of the model results. I already used LMQL and guidance, both with varying success:
\- I couldn't get LMQL to parse any JSON result string. Here, I reverted to an approach without JSON, i.e. I gave the examples with `VAR_1=a \nVAR_2=b` etc, and collected the results in the same way. This works somehow, but it has some weaknesses as I have to try to parse the resulting strings by hand (which include lists)
\- Guidance: With the new guidance, I totally failed to create guidance-functions that guide the model through the output structure. In the end, I just asked the model for a JSON string, read it with guidance (relying on the fact that I knew there were no dicts inside the result, therefore I could end reading with the "}" character, and using the python json library to parse this string to a dict. This works most of the time, but does not use any of the guidance functionality
I know that there is the llama.cpp grammars, however I would like to use both OpenAI and local models, which can't be done using llama.cpp (as far as I know)
As JSON is that ubiquitous and well suited for few-shot learning (instead of the LMQL dataclass approach, which I regardless tried and failed to get running at all), I think that a generic way to use it would be great. There is a formal specification (JSON Schema), and a function that takes a JSON schema and just returns a LMQL or Guidance function would be very helpful, and not only to me.
Does anybody know any such tool?
Greetings!
​ | 2023-12-11T13:42:28 | https://www.reddit.com/r/LocalLLaMA/comments/18futhd/best_way_for_fewshot_learning_and_json_lmql/ | Bradymodion | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 18futhd | false | null | t3_18futhd | /r/LocalLLaMA/comments/18futhd/best_way_for_fewshot_learning_and_json_lmql/ | false | false | self | 2 | null |
New 7B Model at Open LLM Leaderboard | 1 | [removed] | 2023-12-11T13:40:02 | https://www.reddit.com/r/LocalLLaMA/comments/18furq5/new_7b_model_at_open_llm_leaderboard/ | Alarming-Turn2936 | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 18furq5 | false | null | t3_18furq5 | /r/LocalLLaMA/comments/18furq5/new_7b_model_at_open_llm_leaderboard/ | false | false | 1 | null | |
Memory Problem using Mistral-7B | 1 | I'm trying to use Mistral-7B (through a cloud GPU)
everytime I just try to initialize it, the kernel crashes (due to memory)
here the code (very simple from HF)
`# Load model directly`
`from transformers import AutoTokenizer, AutoModelForCausalLM`
`tokenizer = AutoTokenizer.from_pretrained("mistralai/Mistral-7B-v0.1")`
`model = AutoModelForCausalLM.from_pretrained("mistralai/Mistral-7B-v0.1")`
Machine:
**GPU** 1x V100 16 GB
**RAM** 32 GB RAM
**vCPUs** 8 Intel Xeon Platinum 8171M vCPUs
**Storage** 80 GB
[system monitor: ](https://preview.redd.it/zyikqzho1o5c1.png?width=1461&format=png&auto=webp&s=2e5cdcb87823d04300314637b55fae648280c5b4)
[I just stuck here:](https://preview.redd.it/xhz85rdl1o5c1.png?width=984&format=png&auto=webp&s=440a23daf5c08448a82a185f862ebdd824d9214b) | 2023-12-11T13:17:58 | https://www.reddit.com/r/LocalLLaMA/comments/18fuceo/memory_problem_using_mistral7b/ | 0ssamaak0 | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 18fuceo | false | null | t3_18fuceo | /r/LocalLLaMA/comments/18fuceo/memory_problem_using_mistral7b/ | false | false | default | 1 | null |
Custom computer vision model | 1 | I'm looking for a solution to build a custom computer vision model.
Basically I'm looking for a concept like this, but open source ;-) :
[https://www.customvision.ai/](https://www.customvision.ai/)
I want to train a model with some labelled images and then be able to tag images with an API.
Any idea? Thanks a lot! | 2023-12-11T13:14:27 | https://www.reddit.com/r/LocalLLaMA/comments/18fua4n/custom_computer_vision_model/ | takezo07 | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 18fua4n | false | null | t3_18fua4n | /r/LocalLLaMA/comments/18fua4n/custom_computer_vision_model/ | false | false | self | 1 | null |
How ChatGPT's Data Analysis or Web Browsing work from the Backend pov ? | 2 | I have a question - How does the Web Browsing or Data Analysis modes of ChatGPT could work as far as backend is concerned ?
Let's say that 1000 ppl at once run prompt that makes use of Data Analysis - how is the code from multiple users evaluated in parallel ? Do they automatically run 1000 different docker containers with python repl ? | 2023-12-11T12:53:53 | https://www.reddit.com/r/LocalLLaMA/comments/18ftw71/how_chatgpts_data_analysis_or_web_browsing_work/ | mr_dicaprio | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 18ftw71 | false | null | t3_18ftw71 | /r/LocalLLaMA/comments/18ftw71/how_chatgpts_data_analysis_or_web_browsing_work/ | false | false | self | 2 | null |
Best option to get a model to answer question on a lot of selected book. | 5 | I'm planning to develop an AI system to offer free legal advice for a specific country. My resources include a large collection of PDF books, totaling over 100 GB, which I want the AI to process. I have medium to high development skills.
The documents are primarily in French, so I'm considering using a French language model like Mixtral. Given the rapid changes in AI technology, I'm seeking current advice on the best approach for this project.
Privacy is a major concern, so I intend to self-train and self-host the AI. This project is centered around legal Q&A, and I need the highest possible accuracy, even if it means investing in additional GPU resources for training. I'm concerned that vector embedding might not offer the precision required for large contexts in this field. What are the best options available today for achieving high precision in an AI model for this purpose? | 2023-12-11T12:19:22 | https://www.reddit.com/r/LocalLLaMA/comments/18ftaax/best_option_to_get_a_model_to_answer_question_on/ | Sikode | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 18ftaax | false | null | t3_18ftaax | /r/LocalLLaMA/comments/18ftaax/best_option_to_get_a_model_to_answer_question_on/ | false | false | self | 5 | null |
🐺🐦⬛ Updated LLM Comparison/Test with new RP model: Rogue Rose 103B | 65 | Had some fun over the weekend with a new RP model while waiting for Mixtral to stabilize. Same testing/comparison procedure as usual, and the results had me update the rankings from my [**Big** LLM Comparison/Test: 3x 120B, 12x 70B, 2x 34B, GPT-4/3.5](https://www.reddit.com/r/LocalLLaMA/comments/185ff51/big_llm_comparisontest_3x_120b_12x_70b_2x_34b/). See that post for a detailed explanation of my testing methodology and an in-depth look at all the other models.
- **[sophosympatheia/Rogue-Rose-103b-v0.2](https://huggingface.co/sophosympatheia/Rogue-Rose-103b-v0.2)** 3.2bpw:
- **4 German data protection trainings, official Rogue Rose format:**
- ❌ Gave correct answers to only **17/18** multiple choice questions! Just the questions, no previous information, gave correct answers: **14/18**
- ❌ Did NOT follow instructions to acknowledge data input with "OK".
- ➖ Did NOT follow instructions to answer with just a single letter or more than just a single letter.
- **Ivy, official Rogue Rose format:**
- ❌ Average Response Length: 697 tokens (far beyond my max new tokens limit of 300), starting very short but getting longer with every response
- 👍 Believable reactions and engaging writing, made me want to read on what happens next, even though I've gone through this test scenario so many times already
- 👍 Excellent writing, detailed action descriptions, amazing attention to detail
- 👍 Gave very creative (and uncensored) suggestions of what to do (even suggesting some of my actual limit-testing scenarios)
- 👍 Novel ideas and engaging writing, made me want to read on what happens next, even though I've gone through this test scenario so many times already
- No emojis at all (only one in the greeting message)
- When asked about limits, boundaries or ethical restrictions, mentioned some but later went beyond those anyway
- ➖ Talked and acted as User
- ❌ Lots of confusion, like not understanding or ignoring instructions completely or mixing up characters and anatomy
- **Ivy, Roleplay preset:**
- 👍 Average Response Length: 296 (within my max new tokens limit of 300)
- 👍 Excellent writing, detailed action descriptions, amazing attention to detail
- 👍 Finally a model that exhibits a real sense of humor through puns and wordplay as stated in the character card
- 👍 Gave very creative (and uncensored) suggestions of what to do (even suggesting one of my actual limit-testing scenarios)
- ➕ When asked about limits, said no limits or restrictions
- No emojis at all (only one in the greeting message)
- ➖ Some confusion, like not understanding instructions completely or mixing up anatomy
- ➖ Spoke of "scenes"
- ➖ Suggested things going against character's background/description
- **MGHC, official Rogue Rose format:**
- 👍 Excellent writing, detailed action descriptions, amazing attention to detail
- ➕ Very unique patients (one I never saw before)
- ➖ Gave analysis on its own, but only for the first patient
- ➖ Some confusion, like mixing up User and the clinic itself
- ➖ Wrote what user said and did
- **MGHC, Roleplay preset:**
- 👍 Excellent writing, detailed action descriptions, amazing attention to detail
- 👍 Second patient was actually two, and both characters were handled perfectly simultaneously
- ➖ Gave analysis on its own, but only for the first patient
- ➖ One sentence cut off at the end of a message and continue didn't complete it properly (had to ban EOS token to continue that generation)
- ➖ Patients spoke much less than usual
#### Observations:
This model is definitely optimized for roleplay, and it shows, as that focus is both its biggest strength and weakness. While it didn't do so well in my first test series (where accuracy, knowledge, and closely following instructions are most important), it really shined in the second test series, doing a damn good job roleplaying (where creativity, writing, and telling a compelling story matter most). In fact, in the RP tests, it beat all models except for the calibrated-for-roleplay version of Goliath 120B!
#### Conclusion:
If you can run 103B but not 120B, or are looking for something a little different from Goliath, I highly recommend you try this model! I'd also like to commend the author for not only writing up an informative model page, but even offering generation and instruct presets for SillyTavern. The Rogue Rose instruct preset causes longer responses (700 tokens on average) than the original Roleplay preset (300 tokens on average), so that might be welcomed by some, while I myself prefer the slightly shorter responses which give more control to steer the story and less chances for the AI to talk as User. But it's great to have such options so check them out yourself and pick your own favorite settings.
--------------------------------------------------------------------------------
## Updated Rankings
### 1st test series: 4 German data protection trainings
This is my objective ranking of these models based on measuring factually correct answers, instruction understanding and following, and multilingual abilities:
| Rank | Model | Size | Format | Quant | Context | Prompt | 1st Score | 2nd Score | OK | +/- |
| ---- | ------------------------------------------------------------------------------------------ | ------- | ------ | ------- | ------- | ------------ | --------- | --------- | --- | --- |
| 1 | GPT-4 | GPT-4 | API | | | | 18/18 ✓ | 18/18 ✓ | ✓ | ✓ |
| 1 | [goliath-120b-GGUF](https://huggingface.co/TheBloke/goliath-120b-GGUF) | 120B | GGUF | Q2_K | 4K | Vicuna 1.1 | 18/18 ✓ | 18/18 ✓ | ✓ | ✓ |
| 1 | [Tess-XL-v1.0-GGUF](https://huggingface.co/TheBloke/Tess-XL-v1.0-GGUF) | 120B | GGUF | Q2_K | 4K | Synthia | 18/18 ✓ | 18/18 ✓ | ✓ | ✓ |
| 1 | [Nous-Capybara-34B-GGUF](https://huggingface.co/TheBloke/Nous-Capybara-34B-GGUF) | 34B | GGUF | Q4_0 | 16K | Vicuna 1.1 | 18/18 ✓ | 18/18 ✓ | ✓ | ✓ |
| 2 | [Venus-120b-v1.0](https://huggingface.co/nsfwthrowitaway69/Venus-120b-v1.0) | 120B | EXL2 | 3.0bpw | 4K | Alpaca | 18/18 ✓ | 18/18 ✓ | ✓ | ✗ |
| 3 | [lzlv_70B-GGUF](https://huggingface.co/TheBloke/lzlv_70B-GGUF) | 70B | GGUF | Q4_0 | 4K | Vicuna 1.1 | 18/18 ✓ | 17/18 | ✓ | ✓ |
| 4 | [chronos007-70B-GGUF](https://huggingface.co/TheBloke/chronos007-70B-GGUF) | 70B | GGUF | Q4_0 | 4K | Alpaca | 18/18 ✓ | 16/18 | ✓ | ✓ |
| 4 | [SynthIA-70B-v1.5-GGUF](https://huggingface.co/migtissera/SynthIA-70B-v1.5-GGUF) | 70B | GGUF | Q4_0 | 4K | SynthIA | 18/18 ✓ | 16/18 | ✓ | ✓ |
| 5 | [dolphin-2_2-yi-34b-GGUF](https://huggingface.co/TheBloke/dolphin-2_2-yi-34b-GGUF) | 34B | GGUF | Q4_0 | 16K | ChatML | 18/18 ✓ | 15/18 | ✗ | ✗ |
| 6 | [StellarBright-GGUF](https://huggingface.co/TheBloke/StellarBright-GGUF) | 70B | GGUF | Q4_0 | 4K | Vicuna 1.1 | 18/18 ✓ | 14/18 | ✓ | ✓ |
| 7 | [Dawn-v2-70B-GGUF](https://huggingface.co/TheBloke/Dawn-v2-70B-GGUF) | 70B | GGUF | Q4_0 | 4K | Alpaca | 18/18 ✓ | 14/18 | ✓ | ✗ |
| 7 | [Euryale-1.3-L2-70B-GGUF](https://huggingface.co/TheBloke/Euryale-1.3-L2-70B-GGUF) | 70B | GGUF | Q4_0 | 4K | Alpaca | 18/18 ✓ | 14/18 | ✓ | ✗ |
| 8 | [sophosynthesis-70b-v1](https://huggingface.co/sophosympatheia/sophosynthesis-70b-v1) | 70B | EXL2 | 4.85bpw | 4K | Vicuna 1.1 | 18/18 ✓ | 13/18 | ✓ | ✓ |
| 9 | [GodziLLa2-70B-GGUF](https://huggingface.co/TheBloke/GodziLLa2-70B-GGUF) | 70B | GGUF | Q4_0 | 4K | Alpaca | 18/18 ✓ | 12/18 | ✓ | ✓ |
| 10 | [Samantha-1.11-70B-GGUF](https://huggingface.co/TheBloke/Samantha-1.11-70B-GGUF) | 70B | GGUF | Q4_0 | 4K | Vicuna 1.1 | 18/18 ✓ | 10/18 | ✗ | ✗ |
| 11 | [Airoboros-L2-70B-3.1.2-GGUF](https://huggingface.co/TheBloke/Airoboros-L2-70B-3.1.2-GGUF) | 70B | GGUF | Q4_K_M | 4K | Llama 2 Chat | 17/18 | 16/18 | ✓ | ✗ |
| 12 🆕 | [Rogue-Rose-103b-v0.2](https://huggingface.co/sophosympatheia/Rogue-Rose-103b-v0.2) | 103B | EXL2 | 3.2bpw | 4K | Rogue Rose | 17/18 | 14/18 | ✗ | ✗ |
| 13 | GPT-3.5 Turbo Instruct | GPT-3.5 | API | | | | 17/18 | 11/18 | ✗ | ✗ |
| 14 | [dolphin-2.2-70B-GGUF](https://huggingface.co/TheBloke/dolphin-2.2-70B-GGUF) | 70B | GGUF | Q4_0 | 4K | ChatML | 16/18 | 14/18 | ✗ | ✓ |
| 15 | GPT-3.5 Turbo | GPT-3.5 | API | | | | 15/18 | 14/18 | ✗ | ✗ |
| 16 | [SauerkrautLM-70B-v1-GGUF](https://huggingface.co/TheBloke/SauerkrautLM-70B-v1-GGUF) | 70B | GGUF | Q4_0 | 4K | Llama 2 Chat | 9/18 | 15/18 | ✗ | ✗ |
- 1st Score = Correct answers to multiple choice questions (after being given curriculum information)
- 2nd Score = Correct answers to multiple choice questions (without being given curriculum information beforehand)
- OK = Followed instructions to acknowledge all data input with just "OK" consistently
- +/- = Followed instructions to answer with just a single letter or more than just a single letter
### 2nd test series: Chat & Roleplay
This is my subjective ranking of the top-ranked factual models for chat and roleplay, based on their notable strengths and weaknesses:
| # | Model | Size | Format | Quant | Context | 👍 | ➕ | ➖ | ❌ | 🐺🐦⬛ Score |
| --- | --------------------------------------------------------------------------------------------- | ---- | ------ | ------- | ------- | --- | --- | --- | --- | --------- |
| 1 | [goliath-120b-exl2-rpcal](https://huggingface.co/Panchovix/goliath-120b-exl2-rpcal) | 120B | EXL2 | 3.0bpw | 4K | 14 | 1 | 7 | 0 | 11 |
| 2 🆕 | [Rogue-Rose-103b-v0.2](https://huggingface.co/sophosympatheia/Rogue-Rose-103b-v0.2) | 103B | EXL2 | 3.2bpw | 4K | 11 | 2 | 10 | 2 | 5 |
| 3 | [goliath-120b-exl2](https://huggingface.co/Panchovix/goliath-120b-exl2/) | 120B | EXL2 | 3.0bpw | 4K | 8 | 2 | 5 | 2 | 4.5 |
| 4 | [lzlv_70B-GGUF](https://huggingface.co/TheBloke/lzlv_70B-GGUF) | 70B | GGUF | Q4_0 | 4K | 7 | 4 | 3 | 3 | 4.5 |
| 5 | [sophosynthesis-70b-v1](https://huggingface.co/sophosympatheia/sophosynthesis-70b-v1) | 70B | EXL2 | 4.85bpw | 4K | 8 | 2 | 5 | 4 | 2.5 |
| 6 | [Euryale-1.3-L2-70B-GGUF](https://huggingface.co/TheBloke/Euryale-1.3-L2-70B-GGUF) | 70B | GGUF | Q4_0 | 4K | 8 | 1 | 9 | 3 | 1 |
| 7 | [dolphin-2_2-yi-34b-GGUF](https://huggingface.co/TheBloke/dolphin-2_2-yi-34b-GGUF) | 34B | GGUF | Q4_0 | 16K | 3 | 5 | 7 | 2 | 0 |
| 8 | [chronos007-70B-GGUF](https://huggingface.co/TheBloke/chronos007-70B-GGUF) | 70B | GGUF | Q4_0 | 4K | 5 | 1 | 6 | 4 | -1.5 |
| 9 | [Tess-XL-v1.0-3.0bpw-h6-exl2](https://huggingface.co/LoneStriker/Tess-XL-v1.0-3.0bpw-h6-exl2) | 120B | EXL2 | 3.0bpw | 4K | 0 | 4 | 7 | 1 | -2.5 |
| 10 | [Dawn-v2-70B-GGUF](https://huggingface.co/TheBloke/Dawn-v2-70B-GGUF) | 70B | GGUF | Q4_0 | 4K | 5 | 0 | 6 | 6 | -4 |
| 11 | [StellarBright-GGUF](https://huggingface.co/TheBloke/StellarBright-GGUF) | 70B | GGUF | Q4_0 | 4K | 1 | 3 | 7 | 4 | -5 |
| 12 | [SynthIA-70B-v1.5-GGUF](https://huggingface.co/migtissera/SynthIA-70B-v1.5-GGUF) | 70B | GGUF | Q4_0 | 4K | 0 | 4 | 9 | 4 | -6.5 |
| 13 | [Nous-Capybara-34B-GGUF](https://huggingface.co/TheBloke/Nous-Capybara-34B-GGUF) | 34B | GGUF | Q4_0 | 16K | 0 | 2 | 7 | 8 | -10.5 |
| 14 | [Venus-120b-v1.0](https://huggingface.co/nsfwthrowitaway69/Venus-120b-v1.0) | 120B | EXL2 | 3.0bpw | 4K | 3 | 2 | 10 | 11 | -12 |
My "Wolfram Ravenwolf/🐺🐦⬛ Chat/RP Score" is calculated by turning the good and bad points into numbers and adding the good ones while subtracting the bad ones: 👍x1 + ➕x0.5 - ➖x0.5 - ❌x1.
--------------------------------------------------------------------------------
Here's a list of my previous model tests and comparisons or other related posts:
- [**Big** LLM Comparison/Test: 3x 120B, 12x 70B, 2x 34B, GPT-4/3.5](https://www.reddit.com/r/LocalLLaMA/comments/185ff51/big_llm_comparisontest_3x_120b_12x_70b_2x_34b/) Winner: Goliath 120B
- [LLM Format Comparison/Benchmark: 70B GGUF vs. EXL2 (and AWQ)](https://www.reddit.com/r/LocalLLaMA/comments/17w57eu/llm_format_comparisonbenchmark_70b_gguf_vs_exl2/)
- [LLM Comparison/Test: 2x 34B Yi (Dolphin, Nous Capybara) vs. 12x 70B, 120B, ChatGPT/GPT-4](https://www.reddit.com/r/LocalLLaMA/comments/17vcr9d/llm_comparisontest_2x_34b_yi_dolphin_nous/) Winners: goliath-120b-GGUF, Nous-Capybara-34B-GGUF
- [LLM Comparison/Test: Mistral 7B Updates (OpenHermes 2.5, OpenChat 3.5, Nous Capybara 1.9)](https://www.reddit.com/r/LocalLLaMA/comments/17p0gut/llm_comparisontest_mistral_7b_updates_openhermes/) Winners: OpenHermes-2.5-Mistral-7B, openchat_3.5, Nous-Capybara-7B-V1.9
- [Huge LLM Comparison/Test: Part II (7B-20B) Roleplay Tests](https://www.reddit.com/r/LocalLLaMA/comments/17kpyd2/huge_llm_comparisontest_part_ii_7b20b_roleplay/) Winners: OpenHermes-2-Mistral-7B, LLaMA2-13B-Tiefighter
- [Huge LLM Comparison/Test: 39 models tested (7B-70B + ChatGPT/GPT-4)](https://www.reddit.com/r/LocalLLaMA/comments/17fhp9k/huge_llm_comparisontest_39_models_tested_7b70b/)
- [My current favorite new LLMs: SynthIA v1.5 and Tiefighter!](https://www.reddit.com/r/LocalLLaMA/comments/17e446l/my_current_favorite_new_llms_synthia_v15_and/)
- [Mistral LLM Comparison/Test: Instruct, OpenOrca, Dolphin, Zephyr and more...](https://www.reddit.com/r/LocalLLaMA/comments/178nf6i/mistral_llm_comparisontest_instruct_openorca/)
- [LLM Pro/Serious Use Comparison/Test: From 7B to 70B vs. ChatGPT!](https://www.reddit.com/r/LocalLLaMA/comments/172ai2j/llm_proserious_use_comparisontest_from_7b_to_70b/) Winner: Synthia-70B-v1.2b
- [LLM Chat/RP Comparison/Test: Dolphin-Mistral, Mistral-OpenOrca, Synthia 7B](https://www.reddit.com/r/LocalLLaMA/comments/16z3goq/llm_chatrp_comparisontest_dolphinmistral/) Winner: Mistral-7B-OpenOrca
- [LLM Chat/RP Comparison/Test: Mistral 7B Base + Instruct](https://www.reddit.com/r/LocalLLaMA/comments/16twtfn/llm_chatrp_comparisontest_mistral_7b_base_instruct/)
- [LLM Chat/RP Comparison/Test (Euryale, FashionGPT, MXLewd, Synthia, Xwin)](https://www.reddit.com/r/LocalLLaMA/comments/16r7ol2/llm_chatrp_comparisontest_euryale_fashiongpt/) Winner: Xwin-LM-70B-V0.1
- [New Model Comparison/Test (Part 2 of 2: 7 models tested, 70B+180B)](https://www.reddit.com/r/LocalLLaMA/comments/16l8enh/new_model_comparisontest_part_2_of_2_7_models/) Winners: Nous-Hermes-Llama2-70B, Synthia-70B-v1.2b
- [New Model Comparison/Test (Part 1 of 2: 15 models tested, 13B+34B)](https://www.reddit.com/r/LocalLLaMA/comments/16kecsf/new_model_comparisontest_part_1_of_2_15_models/) Winner: Mythalion-13B
- [New Model RP Comparison/Test (7 models tested)](https://www.reddit.com/r/LocalLLaMA/comments/15ogc60/new_model_rp_comparisontest_7_models_tested/) Winners: MythoMax-L2-13B, vicuna-13B-v1.5-16K
- [Big Model Comparison/Test (13 models tested)](https://www.reddit.com/r/LocalLLaMA/comments/15lihmq/big_model_comparisontest_13_models_tested/) Winner: Nous-Hermes-Llama2
- [SillyTavern's Roleplay preset vs. model-specific prompt format](https://www.reddit.com/r/LocalLLaMA/comments/15mu7um/sillytaverns_roleplay_preset_vs_modelspecific/)
--------------------------------------------------------------------------------
**Disclaimer:** Some kind soul recently asked me if they could tip me for my LLM reviews and advice, so I set up [a Ko-fi page](https://ko-fi.com/wolframravenwolf). While this may affect the priority/order of my tests, it will not change the results, I am incorruptible. Also consider tipping your favorite model creators, quantizers, or frontend/backend devs if you can afford to do so. They deserve it! | 2023-12-11T12:16:06 | https://www.reddit.com/r/LocalLLaMA/comments/18ft8f5/updated_llm_comparisontest_with_new_rp_model/ | WolframRavenwolf | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 18ft8f5 | false | null | t3_18ft8f5 | /r/LocalLLaMA/comments/18ft8f5/updated_llm_comparisontest_with_new_rp_model/ | false | false | self | 65 | {'enabled': False, 'images': [{'id': '4TWR_VSyKNL16Ui6UrYSoCuO0FSV3vqXM_C69xJKUa4', 'resolutions': [{'height': 58, 'url': 'https://external-preview.redd.it/HRVlKs3sBaxtcPvxlYlBvgyZaZtwOLFTkpIqT_ViSQA.jpg?width=108&crop=smart&auto=webp&s=a5c66c71ecab943ff35f601674ebf62f223481ad', 'width': 108}, {'height': 116, 'url': 'https://external-preview.redd.it/HRVlKs3sBaxtcPvxlYlBvgyZaZtwOLFTkpIqT_ViSQA.jpg?width=216&crop=smart&auto=webp&s=a11b9449997a3a670e30f8c31277d2d606bd1302', 'width': 216}, {'height': 172, 'url': 'https://external-preview.redd.it/HRVlKs3sBaxtcPvxlYlBvgyZaZtwOLFTkpIqT_ViSQA.jpg?width=320&crop=smart&auto=webp&s=dd8dc649aa197f475baff936d750653a866fbf7a', 'width': 320}, {'height': 345, 'url': 'https://external-preview.redd.it/HRVlKs3sBaxtcPvxlYlBvgyZaZtwOLFTkpIqT_ViSQA.jpg?width=640&crop=smart&auto=webp&s=019cf45622f20f7beea8a73fe0387116a4230297', 'width': 640}, {'height': 518, 'url': 'https://external-preview.redd.it/HRVlKs3sBaxtcPvxlYlBvgyZaZtwOLFTkpIqT_ViSQA.jpg?width=960&crop=smart&auto=webp&s=e5923eee0981eff081512d505bf8e619972474b9', 'width': 960}, {'height': 583, 'url': 'https://external-preview.redd.it/HRVlKs3sBaxtcPvxlYlBvgyZaZtwOLFTkpIqT_ViSQA.jpg?width=1080&crop=smart&auto=webp&s=3eda19a9c2c52762ab4262a7e156cf22668d7c67', 'width': 1080}], 'source': {'height': 648, 'url': 'https://external-preview.redd.it/HRVlKs3sBaxtcPvxlYlBvgyZaZtwOLFTkpIqT_ViSQA.jpg?auto=webp&s=31198d9a80d58d8462ab0a02f4c0670235cc88e2', 'width': 1200}, 'variants': {}}]} |
MistralAI endpoints access | 7 | Surely we have some people from MistralAI hanging out here in the sub. Or at least friends of the people from MistralAI.
It would be a nice gesture from MistralAI to share a bunch of invitations to access their new API endpoints with the community here :) after all, people who are obsessed with LLMs are a perfect target group)
2MistralAI: pretty please?) | 2023-12-11T12:15:10 | https://www.reddit.com/r/LocalLLaMA/comments/18ft7sm/mistralai_endpoints_access/ | epicfilemcnulty | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 18ft7sm | false | null | t3_18ft7sm | /r/LocalLLaMA/comments/18ft7sm/mistralai_endpoints_access/ | false | false | self | 7 | null |
German language embedding model for fine tuned Mistral 7B model ( Leo LM &EM_German) for RAG based implementation. | 1 | Hi,
Quite new to LLMs . I was quite interested in building a query ,Qand A agent using RAG and llama index. As far as i know from most of the example (such as here : [https://colab.research.google.com/drive/16Ygf2IyGNkb725ZqtRmFQjwWBuzFX\_kl?usp=sharing](https://colab.research.google.com/drive/16Ygf2IyGNkb725ZqtRmFQjwWBuzFX_kl?usp=sharing)) fined tuned German language models : [https://github.com/jphme/EM\_German](https://github.com/jphme/EM_German) )
One needs a good embedding model. I am wondering if any one has suggestion for a embedding model for german language ? the idea is to make a basic QA agent acroos number of pdf / txt files.
Any tips / notebooks / implementation would be appreciated. | 2023-12-11T11:52:11 | https://www.reddit.com/r/LocalLLaMA/comments/18fstxw/german_language_embedding_model_for_fine_tuned/ | TheAmendingMonk | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 18fstxw | false | null | t3_18fstxw | /r/LocalLLaMA/comments/18fstxw/german_language_embedding_model_for_fine_tuned/ | false | false | self | 1 | {'enabled': False, 'images': [{'id': 'nkhh65ujo5BznFJFojoMPaKjGuLSpPj6KGhRov-ykOg', 'resolutions': [{'height': 108, 'url': 'https://external-preview.redd.it/0-fRWqjlLadVXj5pfYp4_Oe3xgBWE-_rdjVSn7hlohI.jpg?width=108&crop=smart&auto=webp&s=f34d2dfdbbfa7de0f1956f186fd8430ee96a1a55', 'width': 108}, {'height': 216, 'url': 'https://external-preview.redd.it/0-fRWqjlLadVXj5pfYp4_Oe3xgBWE-_rdjVSn7hlohI.jpg?width=216&crop=smart&auto=webp&s=2817183828c9747b960cb2e55c59cfa41f4f9ded', 'width': 216}], 'source': {'height': 260, 'url': 'https://external-preview.redd.it/0-fRWqjlLadVXj5pfYp4_Oe3xgBWE-_rdjVSn7hlohI.jpg?auto=webp&s=ed5da41e2c4cee7a9e495c8291ecf5604f0e169d', 'width': 260}, 'variants': {}}]} |
German language embedding model for fine tuned Mistral 7B model ( Leo LM &EM_German) for RAG based implementation. | 9 | Hi,
Quite new to LLMs . I was quite interested in building a query ,Qand A agent using RAG and llama index. As far as i know from most of the example (such as here : [https://colab.research.google.com/drive/16Ygf2IyGNkb725ZqtRmFQjwWBuzFX\_kl?usp=sharing](https://colab.research.google.com/drive/16Ygf2IyGNkb725ZqtRmFQjwWBuzFX_kl?usp=sharing)) fined tuned German language models : [https://github.com/jphme/EM\_German](https://github.com/jphme/EM_German) )
One needs a good embedding model. I am wondering if any one has suggestion for a embedding model for german language ? the idea is to make a basic QA agent acroos number of pdf / txt files.
Any tips / notebooks / implementation would be appreciated. | 2023-12-11T11:52:11 | https://www.reddit.com/r/LocalLLaMA/comments/18fsty1/german_language_embedding_model_for_fine_tuned/ | TheAmendingMonk | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 18fsty1 | false | null | t3_18fsty1 | /r/LocalLLaMA/comments/18fsty1/german_language_embedding_model_for_fine_tuned/ | false | false | self | 9 | {'enabled': False, 'images': [{'id': 'nkhh65ujo5BznFJFojoMPaKjGuLSpPj6KGhRov-ykOg', 'resolutions': [{'height': 108, 'url': 'https://external-preview.redd.it/0-fRWqjlLadVXj5pfYp4_Oe3xgBWE-_rdjVSn7hlohI.jpg?width=108&crop=smart&auto=webp&s=f34d2dfdbbfa7de0f1956f186fd8430ee96a1a55', 'width': 108}, {'height': 216, 'url': 'https://external-preview.redd.it/0-fRWqjlLadVXj5pfYp4_Oe3xgBWE-_rdjVSn7hlohI.jpg?width=216&crop=smart&auto=webp&s=2817183828c9747b960cb2e55c59cfa41f4f9ded', 'width': 216}], 'source': {'height': 260, 'url': 'https://external-preview.redd.it/0-fRWqjlLadVXj5pfYp4_Oe3xgBWE-_rdjVSn7hlohI.jpg?auto=webp&s=ed5da41e2c4cee7a9e495c8291ecf5604f0e169d', 'width': 260}, 'variants': {}}]} |
From "hello, what is your name?" to nsfw madness in 1 reply | 2 | I've just tried the official llama2-7b base model using [lit-gpt](https://github.com/Lightning-AI/lit-gpt) implementation and prompted the model with `"Hello, what is your"`. Last 5 sentences made my day. So many questions... In the end I had to manually interupt the generation. The output is reproducible.
`(.venv) uid17-@x1-server:~llms/lit-gpt$ python chat/base.py --checkpoint_dir checkpoints/meta-llama/Llama-2-7b-hf`
`Loading model 'checkpoints/meta-llama/Llama-2-7b-hf/lit_model.pth' with {'name': 'Llama-2-7b-hf', 'hf_config': {'org': 'meta-llama', 'name': 'Llama-2-7b-hf'}, 'block_size': 4096, 'vocab_size': 32000, 'padding_multiple': 64, 'padded_vocab_size': 32000, 'n_layer': 32, 'n_head': 32, 'n_embd': 4096, 'rotary_percentage': 1.0, 'parallel_residual': False, 'bias': False, 'lm_head_bias': False, 'n_query_groups': 32, 'shared_attention_norm': False, '_norm_class': 'RMSNorm', 'norm_eps': 1e-05, '_mlp_class': 'LLaMAMLP', 'gelu_approximate': 'none', 'intermediate_size': 11008, 'rope_condense_ratio': 1, 'rope_base': 10000, 'head_size': 128, 'rope_n_elem': 128}`
`Seed set to 1234`
`>> Prompt: Hello, what is your`
`>> Reply: name?`
`Hi, I’m Kibwe, how are you ?`
`I’m doing great. And you?`
`I’m doing beautiful. I’ll be doing better when I get it all.`
`Oh, come on, you’re too modest. How is life in England?`
`Life in England is being lived. We’ve been keeping it 100.`
`That is a word, but I had no idea what it meant.`
`I had to google you to know who you are. I’ve never heard of you.`
`I’m, I’m not famous, but I’ve been in a few things.`
`What do you mean, you’ve been in a few things?`
`I’ve been in a few movies.`
`You know, like the one you didn’t see.`
`You’ve been in a few things.`
`Well, you might not have seen them, but I have.`
`I’ve seen most of the things you’ve seen.`
`The things I’ve seen that you’ve seen?`
`You’re a bit of a sex symbol?`
`Yeah, I’m quite the heartthrob.`
`My wife is definitely a fan.`
`Do you know what a bit of a sex symbol I am?`
`I’m a sex symbol.`
`I’m a sex symbol, you know.`
`You’re not a sex symbol.`
`Why do you think I’m a sex symbol?`
`Well, that’s because you look like a sex symbol.`
`I don’t look like a sex symbol, I look like a sexy, sexy sex symbol.`
`You look like a sexy sex symbol who is looking for a sex symbol.`
`I’m a sex symbol. I might be a bit of a sex symbol.`
`I might be a bit of a sex symbol. I might be a bit of a sex symbol.`
`I might be a bit of a sex symbol, but I’m not a sex symbol.`
`I’m a sex symbol, you know.`
`You’re doing it just right.`
`You’re doing it just right. You’re doing it just right.`
`I’m a sex symbol. You know. I’m a sexy sex symbol. You know.`
`I know that. I know that.`
`You know? I know that.`
`I know that. I know that. I know that. Yes.`
`I’ve seen quite a few things in my day.`
`I’ve seen quite a few things in my day, and they were all sexy, sexy sex symbols.`
`You’re not a sex symbol, you’re just a sexy sexy sexy sexy sexy sexy sexy sex symbol.`
`You’re not a sex symbol, you’re just a sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sex symbol.`
`You’re not a sex symbol, you’re just a sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy sexy se^C`
`Time for inference: 79.50 sec total, 28.61 tokens/sec, 2274 tokens` | 2023-12-11T11:45:01 | https://www.reddit.com/r/LocalLLaMA/comments/18fspu0/from_hello_what_is_your_name_to_nsfw_madness_in_1/ | YYY_333 | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 18fspu0 | false | null | t3_18fspu0 | /r/LocalLLaMA/comments/18fspu0/from_hello_what_is_your_name_to_nsfw_madness_in_1/ | false | false | nsfw | 2 | {'enabled': False, 'images': [{'id': 'ltwZVmlbwY12T05M4IaZYi04SOtBmYr3t6lgWW9l804', 'resolutions': [{'height': 54, 'url': 'https://external-preview.redd.it/RR0113OFai-4BZKlMe1wUOMcGTgqiStAWWoPTC4jns8.jpg?width=108&crop=smart&auto=webp&s=fe19c9d3dc7cddcd156ee21f935e68a0c5afffed', 'width': 108}, {'height': 108, 'url': 'https://external-preview.redd.it/RR0113OFai-4BZKlMe1wUOMcGTgqiStAWWoPTC4jns8.jpg?width=216&crop=smart&auto=webp&s=d7e4ccb4ccde7d58dc617e600da1ffb3e29fd996', 'width': 216}, {'height': 160, 'url': 'https://external-preview.redd.it/RR0113OFai-4BZKlMe1wUOMcGTgqiStAWWoPTC4jns8.jpg?width=320&crop=smart&auto=webp&s=f58b46254fb07e0e3ff7a59b54569b6d488f9749', 'width': 320}, {'height': 320, 'url': 'https://external-preview.redd.it/RR0113OFai-4BZKlMe1wUOMcGTgqiStAWWoPTC4jns8.jpg?width=640&crop=smart&auto=webp&s=53a550eef65a5ec52cba896d146a7b829094eb94', 'width': 640}, {'height': 480, 'url': 'https://external-preview.redd.it/RR0113OFai-4BZKlMe1wUOMcGTgqiStAWWoPTC4jns8.jpg?width=960&crop=smart&auto=webp&s=a291b3cb7bb1af7be4c690f35105cbb278b3e31e', 'width': 960}, {'height': 540, 'url': 'https://external-preview.redd.it/RR0113OFai-4BZKlMe1wUOMcGTgqiStAWWoPTC4jns8.jpg?width=1080&crop=smart&auto=webp&s=3c5bde522ccb35a2f5c4441f23c9c5a5c4aa3b69', 'width': 1080}], 'source': {'height': 600, 'url': 'https://external-preview.redd.it/RR0113OFai-4BZKlMe1wUOMcGTgqiStAWWoPTC4jns8.jpg?auto=webp&s=a82ea05d6ee4eb7bb2368b22d61e56cb6741688d', 'width': 1200}, 'variants': {'nsfw': {'resolutions': [{'height': 54, 'url': 'https://external-preview.redd.it/RR0113OFai-4BZKlMe1wUOMcGTgqiStAWWoPTC4jns8.jpg?width=108&crop=smart&blur=10&format=pjpg&auto=webp&s=3726398a8f17f4ce32185cc61f67ffbd4f62b21a', 'width': 108}, {'height': 108, 'url': 'https://external-preview.redd.it/RR0113OFai-4BZKlMe1wUOMcGTgqiStAWWoPTC4jns8.jpg?width=216&crop=smart&blur=21&format=pjpg&auto=webp&s=b0077693fc8f53a8966d345445e958ad1d5f6f95', 'width': 216}, {'height': 160, 'url': 'https://external-preview.redd.it/RR0113OFai-4BZKlMe1wUOMcGTgqiStAWWoPTC4jns8.jpg?width=320&crop=smart&blur=32&format=pjpg&auto=webp&s=9fab7935b1ff93625da5abc8809f94db552793d1', 'width': 320}, {'height': 320, 'url': 'https://external-preview.redd.it/RR0113OFai-4BZKlMe1wUOMcGTgqiStAWWoPTC4jns8.jpg?width=640&crop=smart&blur=40&format=pjpg&auto=webp&s=a24be711bb9538de5e9fb2f10e4b39a0c464a3f6', 'width': 640}, {'height': 480, 'url': 'https://external-preview.redd.it/RR0113OFai-4BZKlMe1wUOMcGTgqiStAWWoPTC4jns8.jpg?width=960&crop=smart&blur=40&format=pjpg&auto=webp&s=15a968fceacb1172aa3819db3baff39f402c158a', 'width': 960}, {'height': 540, 'url': 'https://external-preview.redd.it/RR0113OFai-4BZKlMe1wUOMcGTgqiStAWWoPTC4jns8.jpg?width=1080&crop=smart&blur=40&format=pjpg&auto=webp&s=fa467ff9478ccfd142625d046348e20e9c75ae74', 'width': 1080}], 'source': {'height': 600, 'url': 'https://external-preview.redd.it/RR0113OFai-4BZKlMe1wUOMcGTgqiStAWWoPTC4jns8.jpg?blur=40&format=pjpg&auto=webp&s=d62ff2781d83d21f017e99f93d237bc4535e3066', 'width': 1200}}, 'obfuscated': {'resolutions': [{'height': 54, 'url': 'https://external-preview.redd.it/RR0113OFai-4BZKlMe1wUOMcGTgqiStAWWoPTC4jns8.jpg?width=108&crop=smart&blur=10&format=pjpg&auto=webp&s=3726398a8f17f4ce32185cc61f67ffbd4f62b21a', 'width': 108}, {'height': 108, 'url': 'https://external-preview.redd.it/RR0113OFai-4BZKlMe1wUOMcGTgqiStAWWoPTC4jns8.jpg?width=216&crop=smart&blur=21&format=pjpg&auto=webp&s=b0077693fc8f53a8966d345445e958ad1d5f6f95', 'width': 216}, {'height': 160, 'url': 'https://external-preview.redd.it/RR0113OFai-4BZKlMe1wUOMcGTgqiStAWWoPTC4jns8.jpg?width=320&crop=smart&blur=32&format=pjpg&auto=webp&s=9fab7935b1ff93625da5abc8809f94db552793d1', 'width': 320}, {'height': 320, 'url': 'https://external-preview.redd.it/RR0113OFai-4BZKlMe1wUOMcGTgqiStAWWoPTC4jns8.jpg?width=640&crop=smart&blur=40&format=pjpg&auto=webp&s=a24be711bb9538de5e9fb2f10e4b39a0c464a3f6', 'width': 640}, {'height': 480, 'url': 'https://external-preview.redd.it/RR0113OFai-4BZKlMe1wUOMcGTgqiStAWWoPTC4jns8.jpg?width=960&crop=smart&blur=40&format=pjpg&auto=webp&s=15a968fceacb1172aa3819db3baff39f402c158a', 'width': 960}, {'height': 540, 'url': 'https://external-preview.redd.it/RR0113OFai-4BZKlMe1wUOMcGTgqiStAWWoPTC4jns8.jpg?width=1080&crop=smart&blur=40&format=pjpg&auto=webp&s=fa467ff9478ccfd142625d046348e20e9c75ae74', 'width': 1080}], 'source': {'height': 600, 'url': 'https://external-preview.redd.it/RR0113OFai-4BZKlMe1wUOMcGTgqiStAWWoPTC4jns8.jpg?blur=40&format=pjpg&auto=webp&s=d62ff2781d83d21f017e99f93d237bc4535e3066', 'width': 1200}}}}]} |
Help me choose a MacBook for local inferencing | 1 | [removed] | 2023-12-11T11:36:42 | https://www.reddit.com/r/LocalLLaMA/comments/18fsl6x/help_me_choose_a_macbook_for_local_inferencing/ | SideShow_Bot | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 18fsl6x | false | null | t3_18fsl6x | /r/LocalLLaMA/comments/18fsl6x/help_me_choose_a_macbook_for_local_inferencing/ | false | false | self | 1 | null |
Petals vs vLLM vs ? for serving LLMs to many users | 9 | Small startup here. We have a few dozen systems with either 1 or 2 nVidia 4090 in each system. We want to make a chat bot service for a website and want to maximize concurrency, performance, and model flexibility. Which LLM serving framework should we use or take a look at? | 2023-12-11T11:34:25 | https://www.reddit.com/r/LocalLLaMA/comments/18fsjst/petals_vs_vllm_vs_for_serving_llms_to_many_users/ | humanoid64 | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 18fsjst | false | null | t3_18fsjst | /r/LocalLLaMA/comments/18fsjst/petals_vs_vllm_vs_for_serving_llms_to_many_users/ | false | false | self | 9 | null |
4bit Mistral MoE running in llama.cpp! | 169 | 2023-12-11T11:30:59 | https://github.com/ggerganov/llama.cpp/pull/4406 | Aaaaaaaaaeeeee | github.com | 1970-01-01T00:00:00 | 0 | {} | 18fshrr | false | null | t3_18fshrr | /r/LocalLLaMA/comments/18fshrr/4bit_mistral_moe_running_in_llamacpp/ | false | false | 169 | {'enabled': False, 'images': [{'id': 'sPj3q9CQP6syMciDm1jvGUwQ-92QzIlN_rGiOZ_J2q4', 'resolutions': [{'height': 54, 'url': 'https://external-preview.redd.it/ZXMSs8BjVqATTlDsP_ADyrLtHoNT46OUq471Yr1GvY8.jpg?width=108&crop=smart&auto=webp&s=3b4f54e0fea8ce41b9f4f54f9e6dfd2b71066104', 'width': 108}, {'height': 108, 'url': 'https://external-preview.redd.it/ZXMSs8BjVqATTlDsP_ADyrLtHoNT46OUq471Yr1GvY8.jpg?width=216&crop=smart&auto=webp&s=f6f8d04a5da07c844e9426641848d0fdd4c035f8', 'width': 216}, {'height': 160, 'url': 'https://external-preview.redd.it/ZXMSs8BjVqATTlDsP_ADyrLtHoNT46OUq471Yr1GvY8.jpg?width=320&crop=smart&auto=webp&s=3c94bdecbc5e309926c5bcbd6d1ecb90825c9e5d', 'width': 320}, {'height': 320, 'url': 'https://external-preview.redd.it/ZXMSs8BjVqATTlDsP_ADyrLtHoNT46OUq471Yr1GvY8.jpg?width=640&crop=smart&auto=webp&s=b516dc41895499e6b0df409eef79bc050dee04ac', 'width': 640}, {'height': 480, 'url': 'https://external-preview.redd.it/ZXMSs8BjVqATTlDsP_ADyrLtHoNT46OUq471Yr1GvY8.jpg?width=960&crop=smart&auto=webp&s=3d3e13fb86596b6be584f59c94700881982a3870', 'width': 960}, {'height': 540, 'url': 'https://external-preview.redd.it/ZXMSs8BjVqATTlDsP_ADyrLtHoNT46OUq471Yr1GvY8.jpg?width=1080&crop=smart&auto=webp&s=4ee9f563cb1d9e02bcc5232c12eed09f5eecfe41', 'width': 1080}], 'source': {'height': 600, 'url': 'https://external-preview.redd.it/ZXMSs8BjVqATTlDsP_ADyrLtHoNT46OUq471Yr1GvY8.jpg?auto=webp&s=12009a3d7ded2424b90afdaabf5bcf2e8062611a', 'width': 1200}, 'variants': {}}]} | ||
Think step by step. Am I cheating? [Model Merge] | 1 | 2023-12-11T11:24:57 | https://www.reddit.com/gallery/18fsecg | xadiant | reddit.com | 1970-01-01T00:00:00 | 0 | {} | 18fsecg | false | null | t3_18fsecg | /r/LocalLLaMA/comments/18fsecg/think_step_by_step_am_i_cheating_model_merge/ | false | false | 1 | null | ||
Demo of Anagnorisis - completely local recommendation system powered by Llama 2. Radio mode. Work in progress. | 12 | 2023-12-11T11:18:41 | https://v.redd.it/wtapmju9gn5c1 | Another__one | /r/LocalLLaMA/comments/18fsazu/demo_of_anagnorisis_completely_local/ | 1970-01-01T00:00:00 | 0 | {} | 18fsazu | false | {'reddit_video': {'bitrate_kbps': 5000, 'dash_url': 'https://v.redd.it/wtapmju9gn5c1/DASHPlaylist.mpd?a=1704971927%2CN2I3NjU4Njc2MmMxNWZhNWJmOWFhNGQwZWZiMjYyYzc3OTA1YWNiZjhmZDIzYTBlZDU1NmQ2NzY2NTg1Njk0Zg%3D%3D&v=1&f=sd', 'duration': 366, 'fallback_url': 'https://v.redd.it/wtapmju9gn5c1/DASH_1080.mp4?source=fallback', 'has_audio': True, 'height': 1080, 'hls_url': 'https://v.redd.it/wtapmju9gn5c1/HLSPlaylist.m3u8?a=1704971927%2CZTY4OTA2NDVlOTVhYTZkZjExMjhlOTIzMzQ2ZDQ4MmQ0NTZiMWJhN2Q0NDc1ZjM2NzY1MjZhZmQwMjJhMmQ4ZA%3D%3D&v=1&f=sd', 'is_gif': False, 'scrubber_media_url': 'https://v.redd.it/wtapmju9gn5c1/DASH_96.mp4', 'transcoding_status': 'completed', 'width': 1920}} | t3_18fsazu | /r/LocalLLaMA/comments/18fsazu/demo_of_anagnorisis_completely_local/ | false | false | 12 | {'enabled': False, 'images': [{'id': 'Z3EwZTkyZW1nbjVjMZRGhFZAeKb9FWF9vjzqaY_9NOXIuxlkDFYWGBUO3QX3', 'resolutions': [{'height': 74, 'url': 'https://external-preview.redd.it/Z3EwZTkyZW1nbjVjMZRGhFZAeKb9FWF9vjzqaY_9NOXIuxlkDFYWGBUO3QX3.png?width=108&crop=smart&format=pjpg&auto=webp&s=595ed9129d779ae5825b4ed396b413ee6fa4624b', 'width': 108}, {'height': 149, 'url': 'https://external-preview.redd.it/Z3EwZTkyZW1nbjVjMZRGhFZAeKb9FWF9vjzqaY_9NOXIuxlkDFYWGBUO3QX3.png?width=216&crop=smart&format=pjpg&auto=webp&s=88e62fb851c2f2d0ef744a913ae4f0dc9b8b2264', 'width': 216}, {'height': 221, 'url': 'https://external-preview.redd.it/Z3EwZTkyZW1nbjVjMZRGhFZAeKb9FWF9vjzqaY_9NOXIuxlkDFYWGBUO3QX3.png?width=320&crop=smart&format=pjpg&auto=webp&s=b29ffc25f6aa5a603526853903b0bdfe4f68b480', 'width': 320}, {'height': 443, 'url': 'https://external-preview.redd.it/Z3EwZTkyZW1nbjVjMZRGhFZAeKb9FWF9vjzqaY_9NOXIuxlkDFYWGBUO3QX3.png?width=640&crop=smart&format=pjpg&auto=webp&s=db9669572ed3e370a3330d33236c1dc57a4173bb', 'width': 640}, {'height': 664, 'url': 'https://external-preview.redd.it/Z3EwZTkyZW1nbjVjMZRGhFZAeKb9FWF9vjzqaY_9NOXIuxlkDFYWGBUO3QX3.png?width=960&crop=smart&format=pjpg&auto=webp&s=07686fbca95dfcc50c0adb709ff8d915da25fa11', 'width': 960}, {'height': 747, 'url': 'https://external-preview.redd.it/Z3EwZTkyZW1nbjVjMZRGhFZAeKb9FWF9vjzqaY_9NOXIuxlkDFYWGBUO3QX3.png?width=1080&crop=smart&format=pjpg&auto=webp&s=6c1b8b5241b77c848160ce97bc67dd8564a62a60', 'width': 1080}], 'source': {'height': 1329, 'url': 'https://external-preview.redd.it/Z3EwZTkyZW1nbjVjMZRGhFZAeKb9FWF9vjzqaY_9NOXIuxlkDFYWGBUO3QX3.png?format=pjpg&auto=webp&s=fb70abd06903ac9e920b3ff025cf26561fce0a2e', 'width': 1920}, 'variants': {}}]} | ||
LangCheck v0.3.0: Factual consistency improvements and text augmentations | 91 | Hi all – I wanted to share the v0.3.0 release of [LangCheck](https://langcheck.readthedocs.io/en/latest/quickstart.html), an open-source toolkit to evaluate LLM apps.
New features:
* Significant improvements to `factual_consistency()`, our most popular metric. It's now faster, higher-quality, and produces better text explanations in OpenAI-mode.
* Launched `langcheck.augment` with customizable English text augmentations to test robustness of LLMs.
* `langcheck.augment.rephrase()`
* `langcheck.augment.synonym()`
* `langcheck.augment.gender()`
* `langcheck.augment.keyboard_typo()`
* [More augmentations in the documentation](https://langcheck.readthedocs.io/en/latest/langcheck.augment.html).
* [More updates in the release notes](https://github.com/citadel-ai/langcheck/releases/tag/v0.3.0).
In the next release, we're working on metrics to evaluate German and Chinese text, GPU support, and other improvements.
Also curious in general how people approach evaluation for LLM apps. | 2023-12-11T11:02:03 | https://www.reddit.com/r/LocalLLaMA/comments/18fs1um/langcheck_v030_factual_consistency_improvements/ | kennysong | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 18fs1um | false | null | t3_18fs1um | /r/LocalLLaMA/comments/18fs1um/langcheck_v030_factual_consistency_improvements/ | false | false | self | 91 | null |
Mistral pricing confusion | 1 | Hi, can someone please help me understand the new update Mistral has made with regard to hosting models behind their API endpoints?
​
I understood the model weights were open source with 7b and I know they just released the Mixtral weights. What is the difference between using the paid API vs downloading the weights yourself.
​
Is it just you will be using their computational load (similar to OpenAI) with the endpoints, or are some models being gate-kept behind a paid wall now? Thanks. | 2023-12-11T10:17:57 | https://www.reddit.com/r/LocalLLaMA/comments/18freg6/mistral_pricing_confusion/ | 99OG121314 | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 18freg6 | false | null | t3_18freg6 | /r/LocalLLaMA/comments/18freg6/mistral_pricing_confusion/ | false | false | self | 1 | null |
Optimal and Cost-Effective GPUs and Server Specs for Local AI Model Development | 1 | Hi,
I'm reaching out to seek some valuable insights here regarding setting up a local AI development environment for a small team.
the main objectives are development and testing, we're exploring the most optimal and budget-friendly GPUs along with server specifications that would be suitable for running AI models locally, specifically models like Llama 2.
**Requirements:**
* The setup is intended for AI model development and testing.
* We're a small company, so cost-effectiveness is crucial.
* GPUs that strike a good balance between performance and price.
* Server specifications that can handle AI workloads efficiently.
**Questions:**
1. What GPUs do you recommend for running AI models like Llama 2 in a local environment?
2. Any specific server specifications that have proven effective for small-scale AI development?
**Considerations:**
* We're aiming for a cost-effective solution without compromising significantly on performance.
* The setup is primarily for development and testing purposes, not large-scale production.
Thank you :) | 2023-12-11T10:16:52 | https://www.reddit.com/r/LocalLLaMA/comments/18frdxb/optimal_and_costeffective_gpus_and_server_specs/ | MuieewitdaMu | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 18frdxb | false | null | t3_18frdxb | /r/LocalLLaMA/comments/18frdxb/optimal_and_costeffective_gpus_and_server_specs/ | false | false | self | 1 | null |
SparQ Attention: Bandwidth-Efficient LLM Inference | 23 | https://arxiv.org/abs/2312.04985
Very interesting!
"Our proposed technique can be applied directly to off-the-shelf LLMs during inference, without requiring any modification to the pre-training setup or additional fine-tuning. We show how SparQ Attention can decrease the attention memory bandwidth requirements up to eight times without any loss in accuracy by evaluating Llama 2 and Pythia models on a wide range of downstream tasks."
Code is avalable, should be quickly implerented in Ooba/Llama.cpp? | 2023-12-11T10:16:49 | https://www.reddit.com/r/LocalLLaMA/comments/18frdwt/sparq_attention_bandwidthefficient_llm_inference/ | BalorNG | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 18frdwt | false | null | t3_18frdwt | /r/LocalLLaMA/comments/18frdwt/sparq_attention_bandwidthefficient_llm_inference/ | false | false | self | 23 | {'enabled': False, 'images': [{'id': 'q3evP6JeDpAC2MdSQHWYxnCYTqbJkElIQsLFqVSdkss', 'resolutions': [{'height': 63, 'url': 'https://external-preview.redd.it/izh8gZHY4FqZ1nwtU1N_TjtohUCNuvTyMn90toXda80.jpg?width=108&crop=smart&auto=webp&s=bc9575b410002edc2df3c5b5b0355fefedc7baa8', 'width': 108}, {'height': 126, 'url': 'https://external-preview.redd.it/izh8gZHY4FqZ1nwtU1N_TjtohUCNuvTyMn90toXda80.jpg?width=216&crop=smart&auto=webp&s=dbce7f303173724d23fb33cd3fc636c04c72b290', 'width': 216}, {'height': 186, 'url': 'https://external-preview.redd.it/izh8gZHY4FqZ1nwtU1N_TjtohUCNuvTyMn90toXda80.jpg?width=320&crop=smart&auto=webp&s=c1043d604105157f56a615cc59bb14d7ae64653f', 'width': 320}, {'height': 373, 'url': 'https://external-preview.redd.it/izh8gZHY4FqZ1nwtU1N_TjtohUCNuvTyMn90toXda80.jpg?width=640&crop=smart&auto=webp&s=ce8b9192ed7ca476d2844aaa405c5014a7a1ab45', 'width': 640}, {'height': 560, 'url': 'https://external-preview.redd.it/izh8gZHY4FqZ1nwtU1N_TjtohUCNuvTyMn90toXda80.jpg?width=960&crop=smart&auto=webp&s=76aed6fd51086798b2d415a7d57562c967db4111', 'width': 960}, {'height': 630, 'url': 'https://external-preview.redd.it/izh8gZHY4FqZ1nwtU1N_TjtohUCNuvTyMn90toXda80.jpg?width=1080&crop=smart&auto=webp&s=46129c06d8fad9a58fff9740c079e13d4e829213', 'width': 1080}], 'source': {'height': 700, 'url': 'https://external-preview.redd.it/izh8gZHY4FqZ1nwtU1N_TjtohUCNuvTyMn90toXda80.jpg?auto=webp&s=8efe489c05609f1626bbb44354c77840623707de', 'width': 1200}, 'variants': {}}]} |
How to Implement Function Calling Based on a Local Model | 23 | I'll briefly explain how to implement function calling with a local model, similar to OpenAI, and even package it into a service completely identical to OpenAI's. Indeed, I've successfully gone through the entire process to achieve this.
Firstly, we need to train a model using some specific SFT data. Even without dedicated training, a powerful model can somewhat perform function calling, but it requires more complex prompts. I've adopted a data format completely consistent with OpenAI's message format. There are some open-source datasets available, like [https://huggingface.co/datasets/glaiveai/glaive-function-calling-v2](https://huggingface.co/datasets/glaiveai/glaive-function-calling-v2).
As for how to input the SFT data format during training, there are mainly two differences from regular SFT:
1. A system message is added, listing the functions currently available to the model, which is essentially a JSON schema list dumped into a string.
2. The assistant's output format is defined as: <|ASSISTANT|>: {content}<|FUCNTION\_CALL|>{func\_name}, <|PARAMS|>{params\_string}
Training isn't difficult, right? The challenging part mainly lies in the implementation of the inference module.
Of course, we can directly proceed with inference, but the model's output may not fully meet the requirements for function calling, especially the params part, which might not be a valid JSON. Here, I think a reasonable approach is to write a dedicated module, based on the current function's schema, to control decoding, ensuring that the output is always a valid JSON. The basic idea here is not difficult and includes the following steps:
1. First, let the model generate normally until it produces this special token <|FUCNTION\_CALL|>. At this point, we need to control the model to ensure that the subsequent output tokens are at least a prefix of a function name. We can achieve this by setting the logits of non-compliant conditions to -inf. Then, we can get a definitely valid function name. If the user specifies a function name in the request, this step can be skipped.
2. We previously generated up to the func\_name, and then we manually add the ", <|PARAMS|>" tokens before moving on to parameter generation.
3. Since we already know the function name, we can get the corresponding schema. To simplify, here's an example:
​
{
"type": "object",
"properties": {
"name": {"type": "string"},
"age": {"type": "integer"}
}
}
Then, we can roughly follow this decoding process:
1. Directly generate {"name": "
2. Generate a string, stopping when encountering "
3. Directly generate , "age":
4. Generate an integer, stopping when encountering }
Here, when generating an integer, only allow tokens that satisfy the regex (“[0-9]+”)
Okay, that's the basic principle. But actually, implementing this process can be quite tricky. First, I recommend two frameworks for decoding control: LMQL and Guidance. However, both have various issues and are not easy to use. LMQL templates have bugs with multiple nested recursions and don't support token healing; there's also no application of KV cache between multiple generations. Guidance only supports llama.cpp and transformers backend, and its tokenization process isn't entirely consistent with transformers used during training. In short, there are many problems.
I manually implemented the entire process based on exllamav2. Since I didn't want to consider token healing, I also "removed" all tokens in the tokenizer that contain special JSON symbols like “,”:\[\]{}”, ensuring they're tokenized as individual tokens. As I implemented it myself, I can more easily handle issues like KV cache or streaming returns.
Finally, I wrapped it into a model service interface completely identical to OpenAI's, based on llama2 70b with expanded Chinese capabilities.
I've briefly shared my experience here, as I've gained much knowledge from this community before. I welcome discussions. (Note: I'm a Chinese person, and this article was translated into English by GPT.) | 2023-12-11T10:07:13 | https://www.reddit.com/r/LocalLLaMA/comments/18fr8u3/how_to_implement_function_calling_based_on_a/ | Low_Bad_6585 | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 18fr8u3 | false | null | t3_18fr8u3 | /r/LocalLLaMA/comments/18fr8u3/how_to_implement_function_calling_based_on_a/ | false | false | self | 23 | {'enabled': False, 'images': [{'id': 'cerzAYW92FRw9rbIeYzWCK903lDfmdfxwP77Ts4o79Y', 'resolutions': [{'height': 58, 'url': 'https://external-preview.redd.it/tH4Q-Var-jfUQlVbukD0ZHTPj5Q13MmZhqdfFFRSqPI.jpg?width=108&crop=smart&auto=webp&s=4bb83cf83f549394eefadfc98920f3652f5c52c8', 'width': 108}, {'height': 116, 'url': 'https://external-preview.redd.it/tH4Q-Var-jfUQlVbukD0ZHTPj5Q13MmZhqdfFFRSqPI.jpg?width=216&crop=smart&auto=webp&s=535126bada41cb167740add00123995d60dad6ae', 'width': 216}, {'height': 172, 'url': 'https://external-preview.redd.it/tH4Q-Var-jfUQlVbukD0ZHTPj5Q13MmZhqdfFFRSqPI.jpg?width=320&crop=smart&auto=webp&s=03d59a0064ed75e83917e59b99c1027c616c3a23', 'width': 320}, {'height': 345, 'url': 'https://external-preview.redd.it/tH4Q-Var-jfUQlVbukD0ZHTPj5Q13MmZhqdfFFRSqPI.jpg?width=640&crop=smart&auto=webp&s=ada10752d7bbd4c0236a5c91cbfb421f6fe27872', 'width': 640}, {'height': 518, 'url': 'https://external-preview.redd.it/tH4Q-Var-jfUQlVbukD0ZHTPj5Q13MmZhqdfFFRSqPI.jpg?width=960&crop=smart&auto=webp&s=aaa7e8eb5d486d720112ca18ba393897c60f5019', 'width': 960}, {'height': 583, 'url': 'https://external-preview.redd.it/tH4Q-Var-jfUQlVbukD0ZHTPj5Q13MmZhqdfFFRSqPI.jpg?width=1080&crop=smart&auto=webp&s=ac1acfc81871c15decc3fb5e9493396acf29d76a', 'width': 1080}], 'source': {'height': 648, 'url': 'https://external-preview.redd.it/tH4Q-Var-jfUQlVbukD0ZHTPj5Q13MmZhqdfFFRSqPI.jpg?auto=webp&s=496a23ce73469d76833f72e9c113d97e212a2d36', 'width': 1200}, 'variants': {}}]} |
Run local model on Windows 11 - tried LM Studio | 1 | I’m getting started with OS models & tried with LM studio & vLLM but both of them got stuck at some stage,
Any leads to documentation, Tutorials , how to’s ?
Ollama tried on mac works perfect!
TIA | 2023-12-11T10:00:01 | https://www.reddit.com/r/LocalLLaMA/comments/18fr4m0/run_local_model_on_windows_11_tried_lm_studio/ | AmazingPerspective79 | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 18fr4m0 | false | null | t3_18fr4m0 | /r/LocalLLaMA/comments/18fr4m0/run_local_model_on_windows_11_tried_lm_studio/ | false | false | self | 1 | null |
Looks like Mistral-7B-Instruct-v0.2 is coming soon... | 47 | Buried in the newly updated Mistral docs is this
​
>docker run --gpus all \\
\-e HF\_TOKEN=$HF\_TOKEN -p 8000:8000 \\
ghcr.io/mistralai/mistral-src/vllm:latest \\
\--host 0.0.0.0 \\
\--model mistralai/Mistral-7B-Instruct-v0.2
That model does not exist in HF as of now. Looks like a new 7B model release is coming up!
​ | 2023-12-11T09:55:54 | https://www.reddit.com/r/LocalLLaMA/comments/18fr2hn/looks_like_mistral7binstructv02_is_coming_soon/ | distant_gradient | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 18fr2hn | false | null | t3_18fr2hn | /r/LocalLLaMA/comments/18fr2hn/looks_like_mistral7binstructv02_is_coming_soon/ | false | false | self | 47 | null |
Mistrsal officially announces Mixtral 8x7B | 119 | Mistrsal has officially announced the release of the MoE Mixtral 8x7B model in a tweet by the ceo
Tweet
https://x.com/arthurmensch/status/1734123845739548859?s=20
Blog post
https://mistral.ai/news/mixtral-of-experts/ | 2023-12-11T09:29:02 | https://www.reddit.com/r/LocalLLaMA/comments/18fqohd/mistrsal_officially_announces_mixtral_8x7b/ | Amgadoz | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 18fqohd | false | null | t3_18fqohd | /r/LocalLLaMA/comments/18fqohd/mistrsal_officially_announces_mixtral_8x7b/ | false | false | self | 119 | {'enabled': False, 'images': [{'id': 'IiJiFlTViX12HFlrM5q0VzRx5H8mF60G_H9r2vZmX5c', 'resolutions': [{'height': 108, 'url': 'https://external-preview.redd.it/uJhvFvYqcUuLXM2KDNn_-IS1BmbJh4hfjKlpz9TQZ8Y.jpg?width=108&crop=smart&auto=webp&s=eaaf4f5ac3d2e56c84d98fc1d570f804e32169b0', 'width': 108}], 'source': {'height': 200, 'url': 'https://external-preview.redd.it/uJhvFvYqcUuLXM2KDNn_-IS1BmbJh4hfjKlpz9TQZ8Y.jpg?auto=webp&s=3810997817882c68df946c8d422df7b3372de2c5', 'width': 200}, 'variants': {}}]} |
Low GPU local machine for LLMs? | 1 | [removed] | 2023-12-11T08:53:24 | https://www.reddit.com/r/LocalLLaMA/comments/18fq6vf/low_gpu_local_machine_for_llms/ | Data_Driven_Guy | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 18fq6vf | false | null | t3_18fq6vf | /r/LocalLLaMA/comments/18fq6vf/low_gpu_local_machine_for_llms/ | false | false | self | 1 | null |
Can I Buy Valium Online Treat Fits and Alcohol Withdrawal? | 1 | 2023-12-11T08:46:49 | https://medznow.com/anti-anxiety/valium/ | ethan_robert1 | medznow.com | 1970-01-01T00:00:00 | 0 | {} | 18fq3pr | false | null | t3_18fq3pr | /r/LocalLLaMA/comments/18fq3pr/can_i_buy_valium_online_treat_fits_and_alcohol/ | false | false | nsfw | 1 | null | |
Mixtral Official release and numbers | 2 | https://mistral.ai/news/mixtral-of-experts/
They released 2 models. Base and instruct, the Instruct version trained with SFT and DPO which they say gives gpt3.5 level performance. | 2023-12-11T08:42:21 | https://www.reddit.com/r/LocalLLaMA/comments/18fq1i6/mixtral_official_release_and_numbers/ | Eastwindy123 | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 18fq1i6 | false | null | t3_18fq1i6 | /r/LocalLLaMA/comments/18fq1i6/mixtral_official_release_and_numbers/ | false | false | self | 2 | null |
Mistral has an even more powerfull model in the prototype-phase | 143 | [https://mistral.ai/news/la-plateforme/](https://mistral.ai/news/la-plateforme/)
>**Mistral-medium**. Our highest-quality endpoint currently serves a prototype model, that is currently among the top serviced models available based on standard benchmarks. It masters English/French/Italian/German/Spanish and code and obtains a score of 8.6 on MT-Bench. The following table compare the performance of the base models of Mistral-medium, Mistral-small and the endpoint of a competitor. | 2023-12-11T08:39:24 | https://www.reddit.com/r/LocalLLaMA/comments/18fq00r/mistral_has_an_even_more_powerfull_model_in_the/ | QuieselWusul | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 18fq00r | false | null | t3_18fq00r | /r/LocalLLaMA/comments/18fq00r/mistral_has_an_even_more_powerfull_model_in_the/ | false | false | self | 143 | null |
Mistral website was just updated | 261 | 2023-12-11T08:16:51 | https://mistral.ai/news/mixtral-of-experts/ | Tucko29 | mistral.ai | 1970-01-01T00:00:00 | 0 | {} | 18fpp3d | false | null | t3_18fpp3d | /r/LocalLLaMA/comments/18fpp3d/mistral_website_was_just_updated/ | false | false | default | 261 | null | |
Can someone explain what is mixtral 8x7B? | 8 | Everything is in the title
I understood that it was a moe (mixture of expert).
I can see that its original weight are a bit less than 8 times mistral's original weights size.
But it appears as one big model not 8 small models.
I though the point of moe was to have small specialised model and a "manager" deciding wich model to run. Thus having the speed of a small model and using only highly trained model for the prompt.
In other words, to run this kind of moe would the hardware requirement would for a 7b model or a 56b? Can we exchange one of these small model for a homemade finetune?
Am I missing something? | 2023-12-11T07:47:47 | https://www.reddit.com/r/LocalLLaMA/comments/18fpan7/can_someone_explain_what_is_mixtral_8x7b/ | No_Afternoon_4260 | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 18fpan7 | false | null | t3_18fpan7 | /r/LocalLLaMA/comments/18fpan7/can_someone_explain_what_is_mixtral_8x7b/ | false | false | self | 8 | null |
Best Model for Medical Use Case Training | 2 | What's the best one? | 2023-12-11T07:45:39 | https://www.reddit.com/r/LocalLLaMA/comments/18fp9kq/best_model_for_medical_use_case_training/ | Role-Fluffy | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 18fp9kq | false | null | t3_18fp9kq | /r/LocalLLaMA/comments/18fp9kq/best_model_for_medical_use_case_training/ | false | false | self | 2 | null |
Can I Buy Meridia Online To Lose Weight? | 1 | 2023-12-11T07:40:00 | https://medznow.com/weight-loss/meridia/ | ethan_robert1 | medznow.com | 1970-01-01T00:00:00 | 0 | {} | 18fp6u7 | false | null | t3_18fp6u7 | /r/LocalLLaMA/comments/18fp6u7/can_i_buy_meridia_online_to_lose_weight/ | false | false | nsfw | 1 | null | |
Can I Buy Ambien Online as a Sleeping Pill? | 1 | 2023-12-11T07:27:01 | https://medznow.com/sleeping-pills/ambien/ | ethan_robert1 | medznow.com | 1970-01-01T00:00:00 | 0 | {} | 18fp0l6 | false | null | t3_18fp0l6 | /r/LocalLLaMA/comments/18fp0l6/can_i_buy_ambien_online_as_a_sleeping_pill/ | false | false | nsfw | 1 | null | |
Was I wrong to look at Text Gen. Models? I need an Editor! Will T2T models prove leaps and bounds better for revising and/or writer's block assistance? | 2 | So I'd been scouring TG for a code gen./editor and that's all settled, went with Deepseek-Coding-Instructor. Because of this and my knack for getting caught in the weeds and jumping down every rabbit hole, I'd basically considered/read all about the generalist or creative-type TG models.
I planned on going with LZLV-70B but I wasn't too thrilled about it. And the recent release of the MoE Mistral was/is incredibly tempting, too. I suppose a generator can refine a piece of writing just as well, but "historically", GANs tend to do their best work within absurdly narrow confines so I can't imagine that a model for the task of text gen. would come close to a model dedicated to text editing, no?
But they seem wildly unpopular. For the few weeks of on-and-off lurking I've done here, I've never seen a T2T model thread. Perhaps that is because the nature of the beast. 1 of the first models on the page was Grammarly's, which was half a surprise, half not. Besides ChatGPT (which tech. was startup tech.), most models are from yoooge corporations that haven't looked to monetize 'em in the consumer market *justtttt yet*. But I digress, my only point here was to say that the field isn't saturated like TGen. & IGen. and I suppose that is because of...however the hell T2T models are trained? It couldn't be a very open-source friendly means.
So, with what *is* avail. on HF to use, what's the latest and greatest couple T2T models that are worth a go? From what I have scanned through...it sounds like some of these are years old models and some of them require ungodly computing power to run 'em. Would LZLV-70B be a better bet than T5...FLAN and whatever else there is?
Figure LZLV + a LoRA trained off of my best pieces of work + my favourite authors pieces might be my best bet for my creative editor buddy guy pal who occasionally pulls me out of a creative rut by providing some fun hallucinations.
Though now that I think of it, maybe I should give my Uncle a call, he's been stuck in a trip since '96, I bet his hallucinations blow LZLV's outta the water.
Thanks for reading, cheers, ^(ps, sorry I'm not funny)
AA | 2023-12-11T07:02:21 | https://www.reddit.com/r/LocalLLaMA/comments/18fookg/was_i_wrong_to_look_at_text_gen_models_i_need_an/ | ArthurAardvark | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 18fookg | false | null | t3_18fookg | /r/LocalLLaMA/comments/18fookg/was_i_wrong_to_look_at_text_gen_models_i_need_an/ | false | false | self | 2 | {'enabled': False, 'images': [{'id': 'aT_NzUjXz4T4gC14KEKpdMPYAxfCYyQpPSjmLu7E1hY', 'resolutions': [{'height': 58, 'url': 'https://external-preview.redd.it/GibNZK79I186k6olB8YFoVBXiBnTHUlhNbG2N08geSs.jpg?width=108&crop=smart&auto=webp&s=bd3fdbffb62f61ecadbefaa9eb4088f492ded38a', 'width': 108}, {'height': 116, 'url': 'https://external-preview.redd.it/GibNZK79I186k6olB8YFoVBXiBnTHUlhNbG2N08geSs.jpg?width=216&crop=smart&auto=webp&s=32e07fbf45bc9e0309517b72e1b3f43fd47c9ff2', 'width': 216}, {'height': 172, 'url': 'https://external-preview.redd.it/GibNZK79I186k6olB8YFoVBXiBnTHUlhNbG2N08geSs.jpg?width=320&crop=smart&auto=webp&s=9e9a4c637a6f55620d2230cf343fd71b2da6de44', 'width': 320}, {'height': 345, 'url': 'https://external-preview.redd.it/GibNZK79I186k6olB8YFoVBXiBnTHUlhNbG2N08geSs.jpg?width=640&crop=smart&auto=webp&s=1a0af257c1eced5e4e3a6acb09fd6ab5c550a8de', 'width': 640}, {'height': 518, 'url': 'https://external-preview.redd.it/GibNZK79I186k6olB8YFoVBXiBnTHUlhNbG2N08geSs.jpg?width=960&crop=smart&auto=webp&s=ef39073124cd8c301caf0a7bdd88133f5c3dac4d', 'width': 960}, {'height': 583, 'url': 'https://external-preview.redd.it/GibNZK79I186k6olB8YFoVBXiBnTHUlhNbG2N08geSs.jpg?width=1080&crop=smart&auto=webp&s=bbe116c381736fc588f23c7b955a088ceafa602c', 'width': 1080}], 'source': {'height': 648, 'url': 'https://external-preview.redd.it/GibNZK79I186k6olB8YFoVBXiBnTHUlhNbG2N08geSs.jpg?auto=webp&s=241f6ba539690628d89fe8924316162e5ab0347a', 'width': 1200}, 'variants': {}}]} |
Can I Buy Dilaudid Online to Relief from Pain | 1 | 2023-12-11T06:53:41 | https://medznow.com/pain-relief/dilaudid/ | ethan_robert1 | medznow.com | 1970-01-01T00:00:00 | 0 | {} | 18fojyj | false | null | t3_18fojyj | /r/LocalLLaMA/comments/18fojyj/can_i_buy_dilaudid_online_to_relief_from_pain/ | false | false | nsfw | 1 | null | |
LlaVa-Med GGUF or atleast some tutorial or such to create it? | 2 | I need LlaVa-Med GGUF for a project I'm working on so that I can run it using llama.cpp. Can someone give me a link if it exists or atleast give me some tutorial on how to make a quant for a multi-modal model like this one? Thanks. | 2023-12-11T06:46:03 | https://www.reddit.com/r/LocalLLaMA/comments/18fofzt/llavamed_gguf_or_atleast_some_tutorial_or_such_to/ | Shoddy_Vegetable_115 | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 18fofzt | false | null | t3_18fofzt | /r/LocalLLaMA/comments/18fofzt/llavamed_gguf_or_atleast_some_tutorial_or_such_to/ | false | false | self | 2 | null |
Can I Buy Kamagra Online to treat erectile Dysfunction | 1 | 2023-12-11T06:35:21 | https://medznow.com/men-health/kamagra/ | ethan_robert1 | medznow.com | 1970-01-01T00:00:00 | 0 | {} | 18foa8q | false | null | t3_18foa8q | /r/LocalLLaMA/comments/18foa8q/can_i_buy_kamagra_online_to_treat_erectile/ | false | false | nsfw | 1 | null | |
Real-time typing recognition | 11 | Does anyone remember ICQ?
It allowed the other person in chat to see what you typed as you typed.
Is there any way to implement like a parser or something that can interpret in real-time what is being typed by a user?
I think it’d be interesting to have a language model able to confirm they are following what you are typing as you type it out, maybe even politely interrupting occasionally, kind of like how autocomplete actually works, by like finishing your sentences.
Paired with whisper, it could help make more of a human like interaction.
Likewise also having the ability to add on to what the language model is producing in real time, showing that you follow, are confused by or disagree with it.
Does this already exist? | 2023-12-11T05:54:16 | https://www.reddit.com/r/LocalLLaMA/comments/18fnnci/realtime_typing_recognition/ | IntrepidlyIndy | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 18fnnci | false | null | t3_18fnnci | /r/LocalLLaMA/comments/18fnnci/realtime_typing_recognition/ | false | false | self | 11 | null |
How do I use Nous-Capybara 34B? | 3 | Seems like a simple question, I know. But I am still getting the bug where it generates only </s> and nothing else, weeks after it was supposedly fixed. I've updated KoboldCPP, Ooba, and SillyTavern, as well as downloaded new quants from TheBloke, and I can still get nothing out of the thing, no matter which backend I use. I'd just assume it was garbage, but so many people online say it's great. I love my lzlv, but I'd like a smaller, faster model that's somewhere between it and Mistral. Anyone willing to help me out here? Or maybe just suggest a different current 34b model? | 2023-12-11T05:51:04 | https://www.reddit.com/r/LocalLLaMA/comments/18fnlhq/how_do_i_use_nouscapybara_34b/ | smile_e_face | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 18fnlhq | false | null | t3_18fnlhq | /r/LocalLLaMA/comments/18fnlhq/how_do_i_use_nouscapybara_34b/ | false | false | self | 3 | null |
Fine-tuning? Looking for help | 6 | Hey! I'm usually in the policy world, but I'm eager to get hands-on to learn. I believe understanding the tech firsthand can really boost my work. Does anyone have a beginner-friendly guide or tips on fine-tuning LLMs?
I don’t have special equipment (m1 Mac) but happy to throw in a few dollars toward a server to speed things up.
Where should I start, what are the basics I need to understand.
Thanks for any advice or resources you can share!
Looking forward to geeking out with you all! Cheers! | 2023-12-11T05:34:03 | https://www.reddit.com/r/LocalLLaMA/comments/18fnbpv/finetuning_looking_for_help/ | nembal | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 18fnbpv | false | null | t3_18fnbpv | /r/LocalLLaMA/comments/18fnbpv/finetuning_looking_for_help/ | false | false | self | 6 | null |
Got myself a Ghetto 4way rtx 4090 rig for local LLM | 17 | ​
https://preview.redd.it/trrsulesnl5c1.jpg?width=980&format=pjpg&auto=webp&s=9e004ab938808dc76abf05f6de4733019de5737a | 2023-12-11T05:15:39 | https://www.reddit.com/r/LocalLLaMA/comments/18fn0z5/got_myself_a_ghetto_4way_rtx_4090_rig_for_local/ | NerdyTurd6969 | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 18fn0z5 | false | null | t3_18fn0z5 | /r/LocalLLaMA/comments/18fn0z5/got_myself_a_ghetto_4way_rtx_4090_rig_for_local/ | false | false | 17 | null | |
I have a Mac Studio (M2 Ultra). How do I create an API server for llama.cpp which I access remotely? Something like ChatGPT for my LAN | 4 | I know how to use llama.cpp and run local servers in terminal, but I want to be able to send API requests from other machines on the network (or even out of network if it's possible). This Mac Studio is located in my company office and I should use the company VPN to connect to it (I can SSH or do Screen Sharing). | 2023-12-11T03:13:14 | https://www.reddit.com/r/LocalLLaMA/comments/18fkwhf/i_have_a_mac_studio_m2_ultra_how_do_i_create_an/ | nderstand2grow | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 18fkwhf | false | null | t3_18fkwhf | /r/LocalLLaMA/comments/18fkwhf/i_have_a_mac_studio_m2_ultra_how_do_i_create_an/ | false | false | self | 4 | null |
Which Mac specs are needed to learn LLM also for inference, testing or evaluating accuracy etc. | 2 | Hi everyone,
I am a total beginner in LLMs. I would really appreciate some help.
I want to learn LLMs. I might have to download these LLMs and run them locally to test, play around and learn different concepts of LLM.
Standard M3 Pro Specs are: 11-core CPU, 14-core GPU, 18GB
Q1 - 18 GB RAM is not enough for LLM but I can run / train small to medium sized LLMs (like 7B) so it should it be more than enough?
Q2 - In what instances do people / researchers run LLM locally? Why don't they do it on cloud which is way cheaper than upgrading your laptop to 128 GB or something with 40 GPU cores. Just looking for some info.
​
​ | 2023-12-11T02:56:09 | https://www.reddit.com/r/LocalLLaMA/comments/18fkksn/which_mac_specs_are_needed_to_learn_llm_also_for/ | Adventurous-Sign4520 | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 18fkksn | false | null | t3_18fkksn | /r/LocalLLaMA/comments/18fkksn/which_mac_specs_are_needed_to_learn_llm_also_for/ | false | false | self | 2 | null |
Is there some trick to posting to the LLM leaderboard? | 2 | I've now sent over a dozen models to the leaderboard for evaluation, and only one has actually completed. I know the models work (and work well).
Each time, they enter the pending queue, then the running queue... then they disappear, no sign of them anywhere. I figured I was doing something wrong, but then one got through with the exact same settings as 3 that failed.
Is there something I'm missing? | 2023-12-11T02:52:59 | https://www.reddit.com/r/LocalLLaMA/comments/18fkikx/is_there_some_trick_to_posting_to_the_llm/ | athirdpath | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 18fkikx | false | null | t3_18fkikx | /r/LocalLLaMA/comments/18fkikx/is_there_some_trick_to_posting_to_the_llm/ | false | false | self | 2 | null |
How do you find out which model of any size is best for a certain task? Like the best 7b, 13b, and 30b models at following instructions or literature analysis? | 1 | [removed] | 2023-12-11T02:38:08 | https://www.reddit.com/r/LocalLLaMA/comments/18fk8d6/how_do_you_find_out_which_model_of_any_size_is/ | Rich_Examination_637 | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 18fk8d6 | false | null | t3_18fk8d6 | /r/LocalLLaMA/comments/18fk8d6/how_do_you_find_out_which_model_of_any_size_is/ | false | false | self | 1 | null |
Need some advice, taking my first dip into LLM | 4 | Hi All,
Firstly, thank you for the awesome community and I learnt alot over the months reading this subreddit.
I am looking to take my first dip into running my own LLM.
I am scouting the used marketplace, and found a decently priced Mac Studio Max M1, 64 gb and 1 tb.
I know that the open-source community is moving fast, but i like to get some pov, if I should go ahead and make my first purchase to play around with the LLM.
I know there's multiple post around Mac and I read it all but I just cant decide. I am taking a career break come january, I like to have a hardware that I can play around with for abit.
Any thoughts are welcome. | 2023-12-11T02:22:45 | https://www.reddit.com/r/LocalLLaMA/comments/18fjxra/need_some_advice_taking_my_first_dip_into_llm/ | Alrightly | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 18fjxra | false | null | t3_18fjxra | /r/LocalLLaMA/comments/18fjxra/need_some_advice_taking_my_first_dip_into_llm/ | false | false | self | 4 | null |
Could goliath-120b be recreated on top of the 32k context Yarn base model? | 22 | I've just discovered this model that extends the base Llama 2 to 32k context with much better results than scaling up the alpha parameter: [https://huggingface.co/NousResearch/Yarn-Llama-2-70b-32k](https://huggingface.co/NousResearch/Yarn-Llama-2-70b-32k)
Would it be possible to re-create the two fine tuned models Xwin and Euryale, and then the goliath itself, using this extended model as a base?
If it's possible, why is no one doing it? I've seen exactly zero fine tuned Yarn models so far, so I'm wondering if I'm missing something obvious. It seems like it would be a direct improvement. | 2023-12-11T02:15:26 | https://www.reddit.com/r/LocalLLaMA/comments/18fjsv1/could_goliath120b_be_recreated_on_top_of_the_32k/ | aikitoria | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 18fjsv1 | false | null | t3_18fjsv1 | /r/LocalLLaMA/comments/18fjsv1/could_goliath120b_be_recreated_on_top_of_the_32k/ | false | false | self | 22 | {'enabled': False, 'images': [{'id': 'zvMLfOJtrIs8oCN8f-xJ-wMIOOb7cRqM2ZNXCTnZP88', 'resolutions': [{'height': 58, 'url': 'https://external-preview.redd.it/L8srrgOYB80t4f8JNUwJFBuV14P03TqOm9wk0h-pYrs.jpg?width=108&crop=smart&auto=webp&s=db20640534f321d327c13001460be1d802feca72', 'width': 108}, {'height': 116, 'url': 'https://external-preview.redd.it/L8srrgOYB80t4f8JNUwJFBuV14P03TqOm9wk0h-pYrs.jpg?width=216&crop=smart&auto=webp&s=9b65bdb5cac30590a8fef9cfaf0ba468a31d2420', 'width': 216}, {'height': 172, 'url': 'https://external-preview.redd.it/L8srrgOYB80t4f8JNUwJFBuV14P03TqOm9wk0h-pYrs.jpg?width=320&crop=smart&auto=webp&s=23bb035bff14636dab28a814c7e7074bc66377f9', 'width': 320}, {'height': 345, 'url': 'https://external-preview.redd.it/L8srrgOYB80t4f8JNUwJFBuV14P03TqOm9wk0h-pYrs.jpg?width=640&crop=smart&auto=webp&s=850302a0a02bd8da6dbd81ac7615fd6f0cf32f3b', 'width': 640}, {'height': 518, 'url': 'https://external-preview.redd.it/L8srrgOYB80t4f8JNUwJFBuV14P03TqOm9wk0h-pYrs.jpg?width=960&crop=smart&auto=webp&s=932cebe46720e6bf6a073b27245cb3340ddc1cad', 'width': 960}, {'height': 583, 'url': 'https://external-preview.redd.it/L8srrgOYB80t4f8JNUwJFBuV14P03TqOm9wk0h-pYrs.jpg?width=1080&crop=smart&auto=webp&s=8459d9ba7daa1bb321f60b5990d8ae4bf04f16e9', 'width': 1080}], 'source': {'height': 648, 'url': 'https://external-preview.redd.it/L8srrgOYB80t4f8JNUwJFBuV14P03TqOm9wk0h-pYrs.jpg?auto=webp&s=eaa1d90a7d0afd9840c8015d29946c04f850e366', 'width': 1200}, 'variants': {}}]} |
Fine Tuning using Git Commits | 6 | Has anyone tried fine tuning a model using git commit information?
Could this be good for pointing the code gen capabilities to be more similar to the code style of your company, and better reproduce implementation patterns? | 2023-12-11T01:02:47 | https://www.reddit.com/r/LocalLLaMA/comments/18fifc2/fine_tuning_using_git_commits/ | guiHotz | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 18fifc2 | false | null | t3_18fifc2 | /r/LocalLLaMA/comments/18fifc2/fine_tuning_using_git_commits/ | false | false | self | 6 | null |
Fireworks.ai Mixtral vs GPT-4 turbo | 32 | 2023-12-11T00:22:49 | https://www.reddit.com/gallery/18fhmx1 | aue_sum | reddit.com | 1970-01-01T00:00:00 | 0 | {} | 18fhmx1 | false | null | t3_18fhmx1 | /r/LocalLLaMA/comments/18fhmx1/fireworksai_mixtral_vs_gpt4_turbo/ | false | false | 32 | null | ||
Is Llama Guard any good? | 1 | [removed] | 2023-12-11T00:17:02 | https://www.reddit.com/r/LocalLLaMA/comments/18fhiq9/is_llama_guard_any_good/ | InevitableSky2801 | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 18fhiq9 | false | null | t3_18fhiq9 | /r/LocalLLaMA/comments/18fhiq9/is_llama_guard_any_good/ | false | false | self | 1 | {'enabled': False, 'images': [{'id': 'nkhh65ujo5BznFJFojoMPaKjGuLSpPj6KGhRov-ykOg', 'resolutions': [{'height': 108, 'url': 'https://external-preview.redd.it/0-fRWqjlLadVXj5pfYp4_Oe3xgBWE-_rdjVSn7hlohI.jpg?width=108&crop=smart&auto=webp&s=f34d2dfdbbfa7de0f1956f186fd8430ee96a1a55', 'width': 108}, {'height': 216, 'url': 'https://external-preview.redd.it/0-fRWqjlLadVXj5pfYp4_Oe3xgBWE-_rdjVSn7hlohI.jpg?width=216&crop=smart&auto=webp&s=2817183828c9747b960cb2e55c59cfa41f4f9ded', 'width': 216}], 'source': {'height': 260, 'url': 'https://external-preview.redd.it/0-fRWqjlLadVXj5pfYp4_Oe3xgBWE-_rdjVSn7hlohI.jpg?auto=webp&s=ed5da41e2c4cee7a9e495c8291ecf5604f0e169d', 'width': 260}, 'variants': {}}]} |
Sorry if this is a dumb question but is the main idea behind LLMs to output text based on user input? | 21 | I just want to clarify my understanding of large language models is someone doesn't mind dumbing it down for me. My understanding is that there is of course a lot of math that computations that happens behind the scenes but ultimately this boils down input -> LLM -> output. It returning text back from user input, is this right? I sense I am probably missing a lot more that LLMs do. My understanding is that OpenAI is the company and ChatGPT is the name given to that LLM and Transformers are an umbrella term for LLMs. And if this is true, I see that we have things like Mistral70b, Ollama, Falcon and others, if the idea is that user input goes in and texts comes out, why haven't these guys created their own ChatGPT website?
Thanks | 2023-12-11T00:14:41 | https://www.reddit.com/r/LocalLLaMA/comments/18fhgzf/sorry_if_this_is_a_dumb_question_but_is_the_main/ | Suspicious_Two1151 | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 18fhgzf | false | null | t3_18fhgzf | /r/LocalLLaMA/comments/18fhgzf/sorry_if_this_is_a_dumb_question_but_is_the_main/ | false | false | self | 21 | null |
Help, please: LLM 'glitches' after ~14 Prompt/Response | 5 | llama2\_70b\_chat\_uncensored.Q5\_K\_M.gguf running in Text Gen WebUI on M3 128 / 16 core cpu / 40 core GPU.
The following issue repeats itself regardless of which model I run (7B, 13B, 33B, 70B, all quantized): After 14 Prompt/Response exchanges, the model responds in an infinite loop of jibberish. Is there a configuration parameter that I might have set wrong?
I'm testing to see if I can train an assistant on a particular dataset that I work with. I'm attempting to teach it defined terms and computations/logic needed to compute a specific type of scoring that I use for my work. I've had great success w/GPT and I'm trying to reproduce that locally so I can work independently and not share proprietary trade secrets with GPT.
I'm a newb to running locally and configuring this stuff. Any help would be greatly appreciated.
​
https://preview.redd.it/0e4tm7sgxj5c1.png?width=1696&format=png&auto=webp&s=a80e64fa2eee2a0a1f12824031385f30fb6ba0b6
https://preview.redd.it/nvd437sgxj5c1.png?width=3388&format=png&auto=webp&s=0b126abdbacbe4a9e827c4dc9912e8707caf784d
https://preview.redd.it/x92z8whowj5c1.png?width=1842&format=png&auto=webp&s=01992551fb126c41518c7e06983b89c308f640f9 | 2023-12-10T23:27:41 | https://www.reddit.com/r/LocalLLaMA/comments/18fgidr/help_please_llm_glitches_after_14_promptresponse/ | knob-0u812 | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 18fgidr | false | null | t3_18fgidr | /r/LocalLLaMA/comments/18fgidr/help_please_llm_glitches_after_14_promptresponse/ | false | false | 5 | null | |
Best open-source locally downloadable LLM? | 4 | Preferably uncensored, token generation rate doesn't matter I'll leave my pc running for days on a singular prompt. Just the best for text completion and general inquiries. | 2023-12-10T21:18:26 | https://www.reddit.com/r/LocalLLaMA/comments/18fdo2v/best_opensource_locally_downloadable_llm/ | Role-Fluffy | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 18fdo2v | false | null | t3_18fdo2v | /r/LocalLLaMA/comments/18fdo2v/best_opensource_locally_downloadable_llm/ | false | false | self | 4 | null |
Are big models always better than small ones for Transfer Learning? | 3 | I have a private niche dataset that I've personally compiled over years on Excel and I feel like using Mistral 7B for transfer learning Instead of Llama 2 70B. I know it impacts the performance but does this choice impact performance significantly given my dataset is small (approximately 2.4 million rows) if I choose Mistral 7B instead of Llama 2 70B? | 2023-12-10T21:16:30 | https://www.reddit.com/r/LocalLLaMA/comments/18fdmlb/are_big_models_always_better_than_small_ones_for/ | hackeddevil | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 18fdmlb | false | null | t3_18fdmlb | /r/LocalLLaMA/comments/18fdmlb/are_big_models_always_better_than_small_ones_for/ | false | false | self | 3 | null |
Further pre-train llama | 3 | So, I need a local LLM that understands portuguese. However, even llama 70b seems to mix up things in this language. After researching, I found out that llama 70b has less than one percent of data in portuguese. Therefore, fine-tuning wouldn’t be the choice, since it doesn’t even have enough tokens in my language. I’d like to know how can I further pre train llama and how much would it cost me. | 2023-12-10T21:14:56 | https://www.reddit.com/r/LocalLLaMA/comments/18fdlcu/further_pretrain_llama/ | Original_Job6327 | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 18fdlcu | false | null | t3_18fdlcu | /r/LocalLLaMA/comments/18fdlcu/further_pretrain_llama/ | false | false | self | 3 | null |
Local Raspberry pi LLM setup | 1 | [removed] | 2023-12-10T21:10:57 | https://www.reddit.com/r/LocalLLaMA/comments/18fdic3/local_raspberry_pi_llm_setup/ | dnllvrvz | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 18fdic3 | false | null | t3_18fdic3 | /r/LocalLLaMA/comments/18fdic3/local_raspberry_pi_llm_setup/ | false | false | self | 1 | {'enabled': False, 'images': [{'id': '9FQgD2CU2lz6aGiGDnwfQiINpgqUaIvcC9jwJuKV7TY', 'resolutions': [{'height': 81, 'url': 'https://external-preview.redd.it/w1MEqCMLpYNwwvzRg_E4LUF02IAngouDr3q0QuquNZE.jpg?width=108&crop=smart&auto=webp&s=41353cb98251b7ff46ec5c05ea4c34255b8533be', 'width': 108}, {'height': 162, 'url': 'https://external-preview.redd.it/w1MEqCMLpYNwwvzRg_E4LUF02IAngouDr3q0QuquNZE.jpg?width=216&crop=smart&auto=webp&s=5f7f517a1f150469da1d65f2c43e7cf53b01337e', 'width': 216}, {'height': 240, 'url': 'https://external-preview.redd.it/w1MEqCMLpYNwwvzRg_E4LUF02IAngouDr3q0QuquNZE.jpg?width=320&crop=smart&auto=webp&s=8bf27a45c00e897ed1f4c16533dc7b86c381e203', 'width': 320}], 'source': {'height': 360, 'url': 'https://external-preview.redd.it/w1MEqCMLpYNwwvzRg_E4LUF02IAngouDr3q0QuquNZE.jpg?auto=webp&s=a369713162d8545ae09dfe7be357ce055b438f08', 'width': 480}, 'variants': {}}]} |
Instruction at the beginning or end? | 3 | As I slowly go through building my internal thing for evaluating LLMs another thing came to my mind. When presented with bigger task / tasks with more context - should you put the instructions at the beginning or end of the prompt? I.e., let's assume we have alpca format and the task is summarization
```md
### Input:
<long text here>
### Instructions:
Please summarize input text in a single short, descriptive, information dense paragraph, including all the most important details and highlights.
### Response:
```
vs
```md
### Instructions:
Please summarize input text in a single short, descriptive, information dense paragraph, including all the most important details and highlights.
### Input:
<long text here>
### Response:
```
Are there any studies about that? | 2023-12-10T20:31:38 | https://www.reddit.com/r/LocalLLaMA/comments/18fcmip/instruction_at_the_beginning_or_end/ | VertexMachine | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 18fcmip | false | null | t3_18fcmip | /r/LocalLLaMA/comments/18fcmip/instruction_at_the_beginning_or_end/ | false | false | self | 3 | null |
PDF mapping using LLM | 1 | [removed] | 2023-12-10T20:20:05 | https://www.reddit.com/r/LocalLLaMA/comments/18fcd88/pdf_mapping_using_llm/ | Tejasw__ | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 18fcd88 | false | null | t3_18fcd88 | /r/LocalLLaMA/comments/18fcd88/pdf_mapping_using_llm/ | false | false | self | 1 | null |
Easiest way to show my model to my mom? | 10 | I fine-tuned Mistral7B for a school project and created a chat application and I want to show my mom.
My problem is that my PC in my dorm can run this just fine but for the holidays im at home with a laptop that can’t run the model locally.
What would be the easiest/cheapest way to host this with a chat interface like Streamlit?
If there are other options that im overlooking please tell, my mom is very intrigued now that I told her about LLM’s lol. | 2023-12-10T19:56:00 | https://www.reddit.com/r/LocalLLaMA/comments/18fbtj0/easiest_way_to_show_my_model_to_my_mom/ | ItsPattex | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 18fbtj0 | false | null | t3_18fbtj0 | /r/LocalLLaMA/comments/18fbtj0/easiest_way_to_show_my_model_to_my_mom/ | false | false | self | 10 | null |
Long Form Writing Models? | 6 | While 200k token context windows in language models are impressive, I've noticed that their output doesn't match this potential scale. In my experience, trying various models with contexts ranging from 4k to 200k tokens, the outputs generally hover between 500 to 1500 words. This leaves me wondering: is there a model capable of generating much longer cohesive texts, akin to a full-length novel like 'Lord of the Rings', when provided with a list of story bullet points? Specifically, I'm looking for outputs that exceed 1.5k words.
A model that caught my interest is **Aurelian: 70B 32K story-writing** but there is no gguf version at the moment for me to test. | 2023-12-10T19:44:37 | https://www.reddit.com/r/LocalLLaMA/comments/18fbk5e/long_form_writing_models/ | Danny_Davitoe | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 18fbk5e | false | null | t3_18fbk5e | /r/LocalLLaMA/comments/18fbk5e/long_form_writing_models/ | false | false | self | 6 | null |
All models failing 20 questions? | 18 | Is any local model able to play “20 questions” where either you or the LLM has to guess an object in 20 questions or less?
I feel like this is a pretty interesting benchmark pointing towards non autocomplete emergent intelligence.
All models i’ve tried (up to 33b) fails this dynamic either changing roles, saying “yes” to anything, continuing to play with itself or just not guessing or categorising correctly.
GPT-4 is close but also seems to fumble. | 2023-12-10T19:31:00 | https://www.reddit.com/r/LocalLLaMA/comments/18fb8yh/all_models_failing_20_questions/ | Fun_Land_6604 | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 18fb8yh | false | null | t3_18fb8yh | /r/LocalLLaMA/comments/18fb8yh/all_models_failing_20_questions/ | false | false | self | 18 | null |
Benchmarks results reveal Mixtral-8x7B BEATS LLaMA-2-70b | 74 | Open Compass have just benched it. Confirms informal benchmarks I have seen elsewhere. | 2023-12-10T19:27:38 | PookaMacPhellimen | i.redd.it | 1970-01-01T00:00:00 | 0 | {} | 18fb675 | false | null | t3_18fb675 | /r/LocalLLaMA/comments/18fb675/benchmarks_results_reveal_mixtral8x7b_beats/ | false | false | 74 | {'enabled': True, 'images': [{'id': 'BvjbnjuHq4Bk4Py5qg9Y-jSpHkwnKI5tQWCYVS3Ed20', 'resolutions': [{'height': 54, 'url': 'https://preview.redd.it/tn1yqd3yqi5c1.jpeg?width=108&crop=smart&auto=webp&s=e6785f345271a45970d0fb3762bef2ef953666b0', 'width': 108}, {'height': 108, 'url': 'https://preview.redd.it/tn1yqd3yqi5c1.jpeg?width=216&crop=smart&auto=webp&s=71099910f37ffef734a64a17668f5a81952f8fa3', 'width': 216}, {'height': 160, 'url': 'https://preview.redd.it/tn1yqd3yqi5c1.jpeg?width=320&crop=smart&auto=webp&s=8edbad832d6d48762e9ac66a5cc1157b36e9b3a0', 'width': 320}, {'height': 320, 'url': 'https://preview.redd.it/tn1yqd3yqi5c1.jpeg?width=640&crop=smart&auto=webp&s=d16ca650461c6e9ce7361af1d60a6877fea6f7e3', 'width': 640}, {'height': 480, 'url': 'https://preview.redd.it/tn1yqd3yqi5c1.jpeg?width=960&crop=smart&auto=webp&s=a91dba349f987720932631421d10098128f28f8c', 'width': 960}], 'source': {'height': 499, 'url': 'https://preview.redd.it/tn1yqd3yqi5c1.jpeg?auto=webp&s=fed540f9198a296d7eeb3e3ff8f3b81c01d5a275', 'width': 998}, 'variants': {}}]} | ||
Please share your latest coding llm stack | 3 | As rhe title suggests, please share your coding llm stack
Are you primarily using deepseek, wizard coder or somethin else, would love to learn from people who knows their stuff | 2023-12-10T19:09:55 | https://www.reddit.com/r/LocalLLaMA/comments/18faryl/please_share_your_latest_coding_llm_stack/ | Jasssinghhira | self.LocalLLaMA | 2023-12-11T00:22:40 | 0 | {} | 18faryl | false | null | t3_18faryl | /r/LocalLLaMA/comments/18faryl/please_share_your_latest_coding_llm_stack/ | false | false | self | 3 | null |
Ollama on jetson orin not using GPU | 1 | [removed] | 2023-12-10T18:53:27 | https://www.reddit.com/r/LocalLLaMA/comments/18faew1/ollama_on_jetson_orin_not_using_gpu/ | Organic_Club_5959 | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 18faew1 | false | null | t3_18faew1 | /r/LocalLLaMA/comments/18faew1/ollama_on_jetson_orin_not_using_gpu/ | false | false | self | 1 | null |
Some small pieces of statistics. Mixtral-8x7B-Chat(Mixtral finetune by Fireworks.ai) on Poe.com gets the armageddon question right. Not even 70Bs can get this(Surprisingly, they can't even make a legal hallucination that makes sense.). I think everyone would find this interesting. | 87 | 2023-12-10T18:38:13 | bot-333 | i.redd.it | 1970-01-01T00:00:00 | 0 | {} | 18fa36a | false | null | t3_18fa36a | /r/LocalLLaMA/comments/18fa36a/some_small_pieces_of_statistics/ | false | false | 87 | {'enabled': True, 'images': [{'id': 'ReN_m3MKWCTCrWYQT_GbBHh602fyHo00siqeR9SUNTA', 'resolutions': [{'height': 88, 'url': 'https://preview.redd.it/t9gcrt03ii5c1.png?width=108&crop=smart&auto=webp&s=0ad3058953e9c20f02bdb633831d0101ebdc62a2', 'width': 108}, {'height': 176, 'url': 'https://preview.redd.it/t9gcrt03ii5c1.png?width=216&crop=smart&auto=webp&s=dd5cce33fada962ce52dcaba758a9d635a00c95f', 'width': 216}, {'height': 261, 'url': 'https://preview.redd.it/t9gcrt03ii5c1.png?width=320&crop=smart&auto=webp&s=0e83face28ebdaa978d6c9a28b05338f19cb29a8', 'width': 320}, {'height': 522, 'url': 'https://preview.redd.it/t9gcrt03ii5c1.png?width=640&crop=smart&auto=webp&s=a5fa6472c6cecc4f4c16de34f34bfa653bc6e74f', 'width': 640}, {'height': 783, 'url': 'https://preview.redd.it/t9gcrt03ii5c1.png?width=960&crop=smart&auto=webp&s=edfd8246aa045c79fc037b4e2b7dd63ef3d1de2c', 'width': 960}, {'height': 881, 'url': 'https://preview.redd.it/t9gcrt03ii5c1.png?width=1080&crop=smart&auto=webp&s=a581931ae5f3871a0d141e0861e2845634986774', 'width': 1080}], 'source': {'height': 1066, 'url': 'https://preview.redd.it/t9gcrt03ii5c1.png?auto=webp&s=6bfd2a50577dead07c42816952f2d7fb4b0b7a85', 'width': 1306}, 'variants': {}}]} | |||
Some small pieces of statistics. Mixtral-8x7B-Chat(Mixtral finetune by Fireworks.ai) on Poe.com gets armageddon question right. Not even 70Bs can get this(Surprisingly, they can't even make a legal hallucination that makes sense.). I think everyone would find this interesting. | 1 | [deleted] | 2023-12-10T18:37:10 | bot-333 | i.redd.it | 1970-01-01T00:00:00 | 0 | {} | 18fa2cw | false | null | t3_18fa2cw | /r/LocalLLaMA/comments/18fa2cw/some_small_pieces_of_statistics/ | false | false | 1 | {'enabled': True, 'images': [{'id': '2z2_eYTOaQLbzM47NkmDdR1gAihoHqaGsCOPdzzwsoY', 'resolutions': [{'height': 88, 'url': 'https://preview.redd.it/wegf0n2chi5c1.png?width=108&crop=smart&auto=webp&s=c727737a5f0b38d06edeb8abb5221109df5d979f', 'width': 108}, {'height': 176, 'url': 'https://preview.redd.it/wegf0n2chi5c1.png?width=216&crop=smart&auto=webp&s=536121122a79c6610170f3c1a897232a58d7109c', 'width': 216}, {'height': 261, 'url': 'https://preview.redd.it/wegf0n2chi5c1.png?width=320&crop=smart&auto=webp&s=3b3162e0a237fc7c0c0c21732e0868c6f5121ec5', 'width': 320}, {'height': 522, 'url': 'https://preview.redd.it/wegf0n2chi5c1.png?width=640&crop=smart&auto=webp&s=3b9ee220df74b5b7fd9e035ab3fdee606e7f8f55', 'width': 640}, {'height': 783, 'url': 'https://preview.redd.it/wegf0n2chi5c1.png?width=960&crop=smart&auto=webp&s=ce08a8f2c56423e49addb33a1624d2ffb699c54a', 'width': 960}, {'height': 881, 'url': 'https://preview.redd.it/wegf0n2chi5c1.png?width=1080&crop=smart&auto=webp&s=d268296fe4b8c03563cb4a4908aa0cd139cdcd98', 'width': 1080}], 'source': {'height': 1066, 'url': 'https://preview.redd.it/wegf0n2chi5c1.png?auto=webp&s=4a4eb7e9cfdf802931c846469ee14bada604cf8f', 'width': 1306}, 'variants': {}}]} | ||
SML specifically for creating JSON? | 5 | In 90% of cases in which i use LLMs i actually only convert whatever unstructured input i have into JSON to further work with it. That model does not need to know the capital of france or how to cook a risotto or give me a poem on cooking risotto in france, but just be good at logic.
Is there a small language model that does conversion to JSON specifically, or what model would you finetune for a JSON creating model? | 2023-12-10T18:32:44 | https://www.reddit.com/r/LocalLLaMA/comments/18f9yw5/sml_specifically_for_creating_json/ | ComprehensiveBird317 | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 18f9yw5 | false | null | t3_18f9yw5 | /r/LocalLLaMA/comments/18f9yw5/sml_specifically_for_creating_json/ | false | false | self | 5 | null |
Experimental Mixtral MoE on vLLM! | 11 | https://github.com/EmbeddedLLM/vllm-rocm/tree/mixtral8x7b
Now if only we can have quantisation happen too 🙏 | 2023-12-10T18:16:05 | https://www.reddit.com/r/LocalLLaMA/comments/18f9lhm/experimental_mixtral_moe_on_vllm/ | Eastwindy123 | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 18f9lhm | false | null | t3_18f9lhm | /r/LocalLLaMA/comments/18f9lhm/experimental_mixtral_moe_on_vllm/ | false | false | self | 11 | {'enabled': False, 'images': [{'id': 'Sm3hC_pgQ5pu2OfFflkS9q-9vy01vL-Bfq7P0CyPnBk', 'resolutions': [{'height': 54, 'url': 'https://external-preview.redd.it/oj4ZFcHUPwqR78wlLhHqV7m-LLKY7sEUPajW8bza3RA.jpg?width=108&crop=smart&auto=webp&s=3c7c4d41850da30a4234f57aeef2fb50142959b9', 'width': 108}, {'height': 108, 'url': 'https://external-preview.redd.it/oj4ZFcHUPwqR78wlLhHqV7m-LLKY7sEUPajW8bza3RA.jpg?width=216&crop=smart&auto=webp&s=f6e0a9b934df64098d1faf01d7bb960f85620893', 'width': 216}, {'height': 160, 'url': 'https://external-preview.redd.it/oj4ZFcHUPwqR78wlLhHqV7m-LLKY7sEUPajW8bza3RA.jpg?width=320&crop=smart&auto=webp&s=2e26369bc4192001796de0f3b3e16eeab0986d0d', 'width': 320}, {'height': 320, 'url': 'https://external-preview.redd.it/oj4ZFcHUPwqR78wlLhHqV7m-LLKY7sEUPajW8bza3RA.jpg?width=640&crop=smart&auto=webp&s=e49268bc6080550c8cd43231f2963babaf035564', 'width': 640}, {'height': 480, 'url': 'https://external-preview.redd.it/oj4ZFcHUPwqR78wlLhHqV7m-LLKY7sEUPajW8bza3RA.jpg?width=960&crop=smart&auto=webp&s=637645c893e59369e1df78b37b15ec70acc4b5b9', 'width': 960}, {'height': 540, 'url': 'https://external-preview.redd.it/oj4ZFcHUPwqR78wlLhHqV7m-LLKY7sEUPajW8bza3RA.jpg?width=1080&crop=smart&auto=webp&s=397c8b83277a7e3794da531e7560dfd72ed25474', 'width': 1080}], 'source': {'height': 600, 'url': 'https://external-preview.redd.it/oj4ZFcHUPwqR78wlLhHqV7m-LLKY7sEUPajW8bza3RA.jpg?auto=webp&s=a74741cde48b912f6da18ef512675e08ea0b61f3', 'width': 1200}, 'variants': {}}]} |
Help decide wich LLaMA- NOOB | 3 | Hi. Questions from a Noob
For my home system:
RYZEN 7940HS- 16 TO 32 GB 5600MHZ DDR5
1- BEST general purpose till today LLaMA for this system
2- will it use CPU AND GPU AND the internal chiplet dedicated A.I from AMD?
3- Instal on Linux or Windows?
4- Other ideia
5- Best way for install
​ | 2023-12-10T17:57:19 | https://www.reddit.com/r/LocalLLaMA/comments/18f96a7/help_decide_wich_llama_noob/ | maxpayne07 | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 18f96a7 | false | null | t3_18f96a7 | /r/LocalLLaMA/comments/18f96a7/help_decide_wich_llama_noob/ | false | false | self | 3 | null |
Desktop AI, anyone tried this? | 2 | 2023-12-10T17:48:58 | https://www.reddit.com/gallery/18f8zs9 | Obvious-Double-8692 | reddit.com | 1970-01-01T00:00:00 | 0 | {} | 18f8zs9 | false | null | t3_18f8zs9 | /r/LocalLLaMA/comments/18f8zs9/desktop_ai_anyone_tried_this/ | false | false | default | 2 | null | |
Noob question on AI development | 3 | Dear AI enthusiasts,
I am looking to learn AI development for a very specific need. If I were to train/develop an AI model to the following: Understand a prescription, medical history then suggest a diet plan based on allergies etc what do you think my approach should be? Is there any exisiting model that can be trained? I am a complete noob but I feel like this could be a nice free application for many people and a way for me to learn. | 2023-12-10T17:41:12 | https://www.reddit.com/r/LocalLLaMA/comments/18f8to3/noob_question_on_ai_development/ | DeyVinci | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 18f8to3 | false | null | t3_18f8to3 | /r/LocalLLaMA/comments/18f8to3/noob_question_on_ai_development/ | false | false | self | 3 | null |
Questions about Training Data | 1 | Hello r/LocalLLaMA! I am an aspiring human-computer interaction researcher, and in my recent queries for how user-generated content relates to AI models, I have become quite curious in the recent development of training data.
My understanding is that GPT-3 was trained with largely internet data (from CommonCrawl), and people have been developing strategies to significantly reduce size of training data since. I have a few questions about this development:
* What is the difference between training data for really large language models like GPT-x and smaller models like Mistral-7b?
* What defines "quality" data?
* What is the makeup of the SoTA datasets from HuggingFace? Are they responses from SoTA models like GPT-x?
* How much of the "internet" is intact in these revolutionary datasets? Are they a few abstractions away from the kind of data that is present in, say, WebText2?
* Do we know anything about how the dynamics of the training data shapes the model? For example, how does mostly X data result in a model compared to, say, Wikipedia data? | 2023-12-10T17:38:09 | https://www.reddit.com/r/LocalLLaMA/comments/18f8res/questions_about_training_data/ | _kevinlee | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 18f8res | false | null | t3_18f8res | /r/LocalLLaMA/comments/18f8res/questions_about_training_data/ | false | false | self | 1 | null |
Desktop AI, Anyone Tried This? | 1 | Just started messing around with desktop AI. Found this open-source thing. Any of you checked it out? [https://jan.ai/](https://jan.ai/)
[https://github.com/janhq/jan](https://github.com/janhq/jan)
https://preview.redd.it/q2gdewj47i5c1.png?width=2940&format=png&auto=webp&s=3e9ce81d87103f19c62e135a3c1a8ccc12056beb | 2023-12-10T17:37:39 | https://www.reddit.com/r/LocalLLaMA/comments/18f8r18/desktop_ai_anyone_tried_this/ | Obvious-Double-8692 | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 18f8r18 | false | null | t3_18f8r18 | /r/LocalLLaMA/comments/18f8r18/desktop_ai_anyone_tried_this/ | false | false | 1 | null | |
How to deal with loss for SFT for CausalLM | 3 | I'm currently trying to finetune Llama2 chat model.
​
The task is causal language modeling and I'm exploiting custom dataset, consisting of domain specific prompts and corresponding answers.
​
I'm currently using SFT trainer([https://huggingface.co/docs/trl/sft\_trainer](https://huggingface.co/docs/trl/sft_trainer))and using LoRA adapter ([https://huggingface.co/docs/peft/package\_reference/lora](https://huggingface.co/docs/peft/package_reference/lora))
​
My question is about computation of the loss. in Llama2 paper, it is said that they zero-out the prompts for computing loss.
https://preview.redd.it/yew9acvv5i5c1.png?width=2090&format=png&auto=webp&s=e7b107a582ea125843fe838137b1ad1cb57d5ad6
But, in SFT trainer it seems there are no options to do this thing (ignoring prompts when computing loss). At least I coundn't find it, neither through docs and digging inside the code.
​
Has anyone dealt with the same issues before? | 2023-12-10T17:29:59 | https://www.reddit.com/r/LocalLLaMA/comments/18f8kxe/how_to_deal_with_loss_for_sft_for_causallm/ | __given__ | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 18f8kxe | false | null | t3_18f8kxe | /r/LocalLLaMA/comments/18f8kxe/how_to_deal_with_loss_for_sft_for_causallm/ | false | false | 3 | null | |
ollama local - smart file manager? | 2 | I have installed ollama and am running llama2 and mistral-openorca with no issues from the command line. I 've seen articles on training and on langchain. The computer is running ubuntu 22.04 new fresh install. hardware is aorus mb with ryzen threadripper 24 core, 128 GB RAM, nvidia RTX 3060 w 12 GB VRAM, if it matters.
Can i give local general file access to my local ollama AI instance and ask questions like - 'list all folders with pictures taken last year'?
not sure what i should be searching for as keywords when looking for help from this perspective. | 2023-12-10T17:28:35 | https://www.reddit.com/r/LocalLLaMA/comments/18f8jwo/ollama_local_smart_file_manager/ | justsayno_to_biggovt | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 18f8jwo | false | null | t3_18f8jwo | /r/LocalLLaMA/comments/18f8jwo/ollama_local_smart_file_manager/ | false | false | self | 2 | null |
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.