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 ⌀ |
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
Tutorial | Guide: Implementing a Local Language Model-Based Chatbot on Your Project Website | 1 | [removed] | 2024-01-10T14:11:20 | https://www.reddit.com/r/LocalLLaMA/comments/193958j/tutorial_guide_implementing_a_local_language/ | Various_Vermicelli10 | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 193958j | false | null | t3_193958j | /r/LocalLLaMA/comments/193958j/tutorial_guide_implementing_a_local_language/ | false | false | self | 1 | null |
Seeking Advice for Implementing a Local Language Model-Based Chatbot on My Project Website | 1 | [removed] | 2024-01-10T14:08:45 | https://www.reddit.com/r/LocalLLaMA/comments/193939l/seeking_advice_for_implementing_a_local_language/ | Various_Vermicelli10 | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 193939l | false | null | t3_193939l | /r/LocalLLaMA/comments/193939l/seeking_advice_for_implementing_a_local_language/ | false | false | self | 1 | null |
Is this normal - weening off ChatGPT | 23 | For context I went from never using ChatGPT to becoming a super heavy user for my project. I still use it regular to help with coding but I’m starting to get frustrated, even with GPT4, will try copilot soon.
Basically what I’ve experienced is that the more I use it, the more I notice how changes in its back end. For example, GPT4 went from 25 message limit every 3 hour to 40, and I can feel quality drop.
The more I feel this, the more excited I get about exploring local LLMs and learning about the space in general. The feeling of exploration and building something for my own needs is exhilarating, especially in the face of mediocrity. Like, I’d rather spend hours messing around with models and get them to -barely- work, than have a pre-packaged product that just continues to underdeliver.
Do you guys feel this way too? Or like the more you explore the space the more you venture away from ChatGPT? I mean don’t get me wrong, but after the novelty and usage, I find myself wondering “there’s got to be better.”
Disclaimer: This is not a hate post for ChatGPT. | 2024-01-10T13:17:08 | https://www.reddit.com/r/LocalLLaMA/comments/1938247/is_this_normal_weening_off_chatgpt/ | GeeBrain | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 1938247 | false | null | t3_1938247 | /r/LocalLLaMA/comments/1938247/is_this_normal_weening_off_chatgpt/ | false | false | self | 23 | null |
Argilla release a new OpenHermes model outperforming baselines with 54% fewer DPO pairs | 68 | It's a super interesting model that focuses heavily on the data. Via [https://twitter.com/argilla\_io/status/1745057571696693689](https://twitter.com/argilla_io/status/1745057571696693689).
tl;dr
* Fixes issues in [https://huggingface.co/datasets/Intel/orca\_dpo\_pairs](https://huggingface.co/datasets/Intel/orca_dpo_pairs)
* Fine-tunes open-hermes using this newly improved dataset
* Dataset: [https://huggingface.co/datasets/argilla/distilabel-intel-orca-dpo-pairs](https://huggingface.co/datasets/argilla/distilabel-intel-orca-dpo-pairs)
* Model: [https://huggingface.co/argilla/distilabeled-Hermes-2.5-Mistral-7B](https://huggingface.co/argilla/distilabeled-Hermes-2.5-Mistral-7B)
Argilla uses their [distilabel](https://distilabel.argilla.io/latest/) framework to fix issues in the orca\_dpo\_pairs dataset. They did this with a couple of hours of work. I think more community effort around dataset fixes will be super impactful, IMO. | 2024-01-10T13:12:26 | https://www.reddit.com/r/LocalLLaMA/comments/1937ywc/argilla_release_a_new_openhermes_model/ | dvanstrien | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 1937ywc | false | null | t3_1937ywc | /r/LocalLLaMA/comments/1937ywc/argilla_release_a_new_openhermes_model/ | false | false | self | 68 | {'enabled': False, 'images': [{'id': '7NaGS8I38nAU0nNKTNS_nn_GD6EE4LmOqoFGU-qg5DA', 'resolutions': [{'height': 108, 'url': 'https://external-preview.redd.it/f69canj2f_SciJ_S7In8D4TX3KjMkl2W2uanMK5kgmQ.jpg?width=108&crop=smart&auto=webp&s=7a3b6eca11d27a06f5c6da4b1ee7f9610775d731', 'width': 108}], 'source': {'height': 140, 'url': 'https://external-preview.redd.it/f69canj2f_SciJ_S7In8D4TX3KjMkl2W2uanMK5kgmQ.jpg?auto=webp&s=37f80498823dd9c5a0424e7afd31b463fe07b438', 'width': 140}, 'variants': {}}]} |
Prompt Management for Rivet | 3 | Just released the prompt management in my Fast API backend for Rivet, enabling to load prompts stored in Yaml files, and transfer them from one project to the other!
Github:[https://github.com/gabrielolympie/VisualAIFlows](https://github.com/gabrielolympie/VisualAIFlows)
​
https://preview.redd.it/gflg3j5w3mbc1.png?width=1608&format=png&auto=webp&s=0f7ad4fb36f8026a955ad2fc17252541951893e2 | 2024-01-10T13:10:37 | https://www.reddit.com/r/LocalLLaMA/comments/1937xns/prompt_management_for_rivet/ | AdventurousSwim1312 | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 1937xns | false | null | t3_1937xns | /r/LocalLLaMA/comments/1937xns/prompt_management_for_rivet/ | false | false | 3 | {'enabled': False, 'images': [{'id': '-43MSuiHaCS2gcZi8gmvZYGznbucIOZvAHJG3LK7pPc', 'resolutions': [{'height': 54, 'url': 'https://external-preview.redd.it/hbeLrRENQX5vIC2PPs2nUec40DT9ZOQ2GTSf29mkWqE.jpg?width=108&crop=smart&auto=webp&s=3c633dcf7f9c6dbb8326abf2b92f5dab8cdf035d', 'width': 108}, {'height': 108, 'url': 'https://external-preview.redd.it/hbeLrRENQX5vIC2PPs2nUec40DT9ZOQ2GTSf29mkWqE.jpg?width=216&crop=smart&auto=webp&s=089139e5100f3bf61b3d66b756a1bcb94b968542', 'width': 216}, {'height': 160, 'url': 'https://external-preview.redd.it/hbeLrRENQX5vIC2PPs2nUec40DT9ZOQ2GTSf29mkWqE.jpg?width=320&crop=smart&auto=webp&s=1f586818d79115ae02e80686df5181c7e4d17296', 'width': 320}, {'height': 320, 'url': 'https://external-preview.redd.it/hbeLrRENQX5vIC2PPs2nUec40DT9ZOQ2GTSf29mkWqE.jpg?width=640&crop=smart&auto=webp&s=977b5719258074659b4f84df4f6169556ef61247', 'width': 640}, {'height': 480, 'url': 'https://external-preview.redd.it/hbeLrRENQX5vIC2PPs2nUec40DT9ZOQ2GTSf29mkWqE.jpg?width=960&crop=smart&auto=webp&s=1755ba322167ad9fccde540b5986e367da77cffa', 'width': 960}, {'height': 540, 'url': 'https://external-preview.redd.it/hbeLrRENQX5vIC2PPs2nUec40DT9ZOQ2GTSf29mkWqE.jpg?width=1080&crop=smart&auto=webp&s=414c7acf0e9d6203a7d701c20c8d60fc98a98c1d', 'width': 1080}], 'source': {'height': 600, 'url': 'https://external-preview.redd.it/hbeLrRENQX5vIC2PPs2nUec40DT9ZOQ2GTSf29mkWqE.jpg?auto=webp&s=dc99f5be1541faf1fb077f76cb4a96ae5f066fb5', 'width': 1200}, 'variants': {}}]} | |
Is only VRAM important for local LLMs? | 3 | Hey!
I currenty have a 9900k + 2080ti PC and a 32gb M1 Pro macbook. Planning to set up a local LLM on the pro m1 for the time being, but the more i get into AI the more i want to plan something more suited / faster for LLM.
Is it just VRAM that matters? So for instance if I'm to pay up additional 1500$ for more Mac studio cores - should I do that or should I get more ram? (effectively making available ram up to 188gb). Or maybe is it better to have 5x 3090? | 2024-01-10T12:43:11 | https://www.reddit.com/r/LocalLLaMA/comments/1937fax/is_only_vram_important_for_local_llms/ | Rabus | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 1937fax | false | null | t3_1937fax | /r/LocalLLaMA/comments/1937fax/is_only_vram_important_for_local_llms/ | false | false | self | 3 | null |
People are getting sick of GPT4 and switching to local LLMs | 309 | 2024-01-10T12:11:25 | fremenmuaddib | i.redd.it | 1970-01-01T00:00:00 | 0 | {} | 1936vm8 | false | null | t3_1936vm8 | /r/LocalLLaMA/comments/1936vm8/people_are_getting_sick_of_gpt4_and_switching_to/ | false | false | 309 | {'enabled': True, 'images': [{'id': 'XfaqkH4UsJtkX2AfLAgtt-RJe38aEifE8EDdMR4XEac', 'resolutions': [{'height': 60, 'url': 'https://preview.redd.it/hxe385d3tlbc1.png?width=108&crop=smart&auto=webp&s=1fc8ca35cf6f6742060c5dc9f1243d2bf9b4e64f', 'width': 108}, {'height': 121, 'url': 'https://preview.redd.it/hxe385d3tlbc1.png?width=216&crop=smart&auto=webp&s=4169e1be15d961f545da07b8eccd7fc8e30df79a', 'width': 216}, {'height': 180, 'url': 'https://preview.redd.it/hxe385d3tlbc1.png?width=320&crop=smart&auto=webp&s=62acadaa1239467af0b9570552f1818d2c1e5e41', 'width': 320}, {'height': 360, 'url': 'https://preview.redd.it/hxe385d3tlbc1.png?width=640&crop=smart&auto=webp&s=d63ea09e166627ad9b7c2c5e62476d1bf3544c5b', 'width': 640}, {'height': 541, 'url': 'https://preview.redd.it/hxe385d3tlbc1.png?width=960&crop=smart&auto=webp&s=ac72e4db8923f481a13563b6dc80731ac3e49c2f', 'width': 960}, {'height': 609, 'url': 'https://preview.redd.it/hxe385d3tlbc1.png?width=1080&crop=smart&auto=webp&s=cabf10ab529b74c4129b17cd1a260982a3e35256', 'width': 1080}], 'source': {'height': 1080, 'url': 'https://preview.redd.it/hxe385d3tlbc1.png?auto=webp&s=9732011f4d52d8f75526d2b95b91d72bf301b2c6', 'width': 1915}, 'variants': {}}]} | |||
Can you bring Q4 model to its FP16/FP32 performance levels? | 8 | With recent releases like TinyLLama 1.1B, which competes with Llama7B, use a "teacher" model to train a much smaller model. Similarly, SDXL's turbo version of stable diffusion adopted a comparable approach.
This made me wonder: is it feasible to train a quantized model using the original model to rectify inaccuracies in the quantized version?
Theoretically, the dataset could be infinite, as any set of tokens can be run on the parent model to fine-tune the 4Q model.
​ | 2024-01-10T12:05:40 | https://www.reddit.com/r/LocalLLaMA/comments/1936s3x/can_you_bring_q4_model_to_its_fp16fp32/ | niftylius | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 1936s3x | false | null | t3_1936s3x | /r/LocalLLaMA/comments/1936s3x/can_you_bring_q4_model_to_its_fp16fp32/ | false | false | self | 8 | null |
Wait for M3 Ultra, M2 Ultra or 2x4090? | 18 | I have to admit, I'm a bit disappointed with the newest 40xx super generation. I still hoped for more vram despite the leaks were already telling us that the new super GPUs would have max 16Gb vram.
So under the light of this release, I'd love to hear your opinions/reviews for buying a capable LLM-Rig.
I initially wanted to go with 2 4090s to be able to load even whole quantisized 70B models into the vram. I heard 2 3090 would be also sufficient, but by the time I would spend this amount of money, I'd rather go with some gpus that will be supported for a bit longer, are faster in inference, and still have warranty. (Yes, I'm aware of the huge price and wattage difference)
The other possible nvidia variant would be a rtx6000 or rtx6000ada but those seem to be quite overpriced, if available at all.
An alternative would be a m2 ultra or the upcoming m3 ultra. Which provide enough unified memory but seem to lack in compability, have slower t/s and especially(!) time to first token.
I'd like to do inference with 70B models, train loras (if possible with the amount of vram/with the m2) and maybe use it for some stable diffusion.
I'm still in favour of the 2 4090s because I guess we will see smarter models with less demand in resource in the future, so a faster time to the first token and t/s might be more valuable than the additional ram of the m(x) ultra.
Which setup do you prefer? Which is the most future proof (if one can tell at all)? Which would you recommend? | 2024-01-10T12:02:29 | https://www.reddit.com/r/LocalLLaMA/comments/1936q5s/wait_for_m3_ultra_m2_ultra_or_2x4090/ | No-Dot-6573 | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 1936q5s | false | null | t3_1936q5s | /r/LocalLLaMA/comments/1936q5s/wait_for_m3_ultra_m2_ultra_or_2x4090/ | false | false | self | 18 | null |
Is it possible to query the number of layers in a (gguf) model with a command line tool? | 1 | I finally got around to trying Mixtral 8x7B yesterday (my jaw is still on the floor) and it would have been nice to know how many layers it had before running it for the first time so I could estimate how many to offload to the GPU.
While it's possible to Google this kind of model specific information, I've had the issue often enough with different models that a command line tool which shows this (and other) info would be valuable.
koboldcpp, llamacpp, etc. do show this information when loading a model, but where the model is a bit large for the system, it takes a while. | 2024-01-10T12:02:06 | https://www.reddit.com/r/LocalLLaMA/comments/1936pwm/is_it_possible_to_query_the_number_of_layers_in_a/ | metamec | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 1936pwm | false | null | t3_1936pwm | /r/LocalLLaMA/comments/1936pwm/is_it_possible_to_query_the_number_of_layers_in_a/ | false | false | self | 1 | null |
Linux vs windows | 4 | I was wondering how much of an performance increase is to be expected when changing from win 11 to Linux. I can only run on cpu. | 2024-01-10T11:39:48 | https://www.reddit.com/r/LocalLLaMA/comments/1936cd0/linux_vs_windows/ | Single-Cow-5163 | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 1936cd0 | false | null | t3_1936cd0 | /r/LocalLLaMA/comments/1936cd0/linux_vs_windows/ | false | false | self | 4 | null |
Literally my first conversation with it | 508 | I wonder how this got triggered | 2024-01-10T11:29:21 | alymahryn | i.redd.it | 1970-01-01T00:00:00 | 0 | {} | 19366g7 | false | null | t3_19366g7 | /r/LocalLLaMA/comments/19366g7/literally_my_first_conversation_with_it/ | false | false | 508 | {'enabled': True, 'images': [{'id': 'DQkfpBs7W0Lwc_mfNqWioBuSZaGQwwBVKUUqU2ERwsM', 'resolutions': [{'height': 73, 'url': 'https://preview.redd.it/hdxx0xbvllbc1.jpeg?width=108&crop=smart&auto=webp&s=9bf1e306aa12cd9537cd5afe3f8f2219f955a6d0', 'width': 108}, {'height': 147, 'url': 'https://preview.redd.it/hdxx0xbvllbc1.jpeg?width=216&crop=smart&auto=webp&s=e185fd4f3348e2f9bb04509740f0264cfd3d587e', 'width': 216}, {'height': 219, 'url': 'https://preview.redd.it/hdxx0xbvllbc1.jpeg?width=320&crop=smart&auto=webp&s=71598eb80622f67a484f3570a15b21739c5fd548', 'width': 320}, {'height': 438, 'url': 'https://preview.redd.it/hdxx0xbvllbc1.jpeg?width=640&crop=smart&auto=webp&s=25709db101781f172fcb678310e80fae6f32dcb3', 'width': 640}, {'height': 657, 'url': 'https://preview.redd.it/hdxx0xbvllbc1.jpeg?width=960&crop=smart&auto=webp&s=6d390fbc27abfe7cc60b425cf70fd76aef34b5a5', 'width': 960}, {'height': 739, 'url': 'https://preview.redd.it/hdxx0xbvllbc1.jpeg?width=1080&crop=smart&auto=webp&s=d478839323faa27a8d3453d61607a1259f59f1c3', 'width': 1080}], 'source': {'height': 2400, 'url': 'https://preview.redd.it/hdxx0xbvllbc1.jpeg?auto=webp&s=2881acdd7d1b390f275104ac392c26f610978045', 'width': 3504}, 'variants': {}}]} | ||
Any good alternatives to RVC for AI cover song maker | 1 | For context, I haven't use RVC yet!! The problem is the install file from the official GitHub page got a **malware** according my antivirus (Bitdiffender) and also the project has been pretty much abandoned!! | 2024-01-10T11:22:59 | https://www.reddit.com/r/LocalLLaMA/comments/19362x1/any_good_alternatives_to_rvc_for_ai_cover_song/ | qualaric | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 19362x1 | false | null | t3_19362x1 | /r/LocalLLaMA/comments/19362x1/any_good_alternatives_to_rvc_for_ai_cover_song/ | false | false | default | 1 | null |
Can I run Llama on +10-year old hardware? | 1 | Hello! I wasn't thinking in running Llama, but I found a plug-in that's part of my note-taking app, Obsidian, which allows Co-piloting with an AI on the cloud AND locally with Llama!
With that being said, I just want to run it for a bit to see how it does.
The problem is that I have a 6-core AMD FX-6300. Overclocked to 4.2ghz on every core.
My GPU is an AMD R9 280 as well. With 3 VRAM.
I have 16gb dual channel DDR3 ram, Which seems to run up to, 2000mhz. Max Bandwidth, PC3-12800 (800Mhz).
Which llama model can I run to inspect only my Obsidian notes, which contain .MD text files? Sometimes there's images or tables / charts.
Posting an image to prove I'm not trolling:
https://preview.redd.it/eg691w7tglbc1.png?width=499&format=png&auto=webp&s=f6bd7a0686bcd946ee586b7bde10376d06e89357 | 2024-01-10T11:02:24 | https://www.reddit.com/r/LocalLLaMA/comments/1935rgg/can_i_run_llama_on_10year_old_hardware/ | Trainer_Red99 | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 1935rgg | false | null | t3_1935rgg | /r/LocalLLaMA/comments/1935rgg/can_i_run_llama_on_10year_old_hardware/ | false | false | 1 | null | |
Is the finetuning time of 67 hours valid for the following configuration? | 1 | I am taking a large text file and fitting them into 19k rows. Each row is exactly 4096 tokens (not characters), which is the maximum context length of llama 2.
GPU: 80GB
My other configurations are
Batch size: 4
Peft method: Qlora
Target modules: q_proj, v_proj
Quantization: 4 bit
Epochs: 1
Rank: 1
Grad accum steps: 4
Learning rate: 0.0001
Optimiser: adamw_apex_fused
Resultant total steps: 19196
Estimated time: 56 hours
GPU memory occupied: 57 GB
Any tips to further reduce the fine tuning time? Increasing batchsize even by 1 is giving OOM. | 2024-01-10T10:32:28 | https://www.reddit.com/r/LocalLLaMA/comments/1935bhq/is_the_finetuning_time_of_67_hours_valid_for_the/ | dire_wolf_cookie | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 1935bhq | false | null | t3_1935bhq | /r/LocalLLaMA/comments/1935bhq/is_the_finetuning_time_of_67_hours_valid_for_the/ | false | false | self | 1 | null |
Local Chat with PDF - Hardware Question | 2 | Hey Guys,
I'm looking for advice regarding a "local Chat with PDF" Laptop on a budget.
I've done some testing using PrivateGPT with 7B Models on my desktop PC and found that generating an answer with my middle-class CPU takes around 40–80 seconds, while GPU accelerated on my RTX 4070 takes only \~ 5 seconds.
Now I wonder if it would be worth it to run a larger 30b Model on a High-End CPU with 64GB of RAM or run the 7b models on a GPU like the RTX 4070 mobile?
VRAM limitation in laptops is quite tough.
What would you go for? How big is the Performance improvement from a higher parameter model?
| 2024-01-10T10:11:39 | https://www.reddit.com/r/LocalLLaMA/comments/19350ew/local_chat_with_pdf_hardware_question/ | Failiiix | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 19350ew | false | null | t3_19350ew | /r/LocalLLaMA/comments/19350ew/local_chat_with_pdf_hardware_question/ | false | false | self | 2 | null |
non ASCII alphabet | 1 | any LLMs that be trained on strange alphabets like the Russian? i need to learn the language
Regards | 2024-01-10T10:04:03 | https://www.reddit.com/r/LocalLLaMA/comments/1934wd8/non_ascii_alphabet/ | Dyonizius | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 1934wd8 | false | null | t3_1934wd8 | /r/LocalLLaMA/comments/1934wd8/non_ascii_alphabet/ | false | false | self | 1 | null |
S-Lora available now with OpenAI v1 Completions API in Lorax! | 19 | 2024-01-10T09:24:26 | https://github.com/predibase/lorax/pull/170 | phoneixAdi | github.com | 1970-01-01T00:00:00 | 0 | {} | 1934c3u | false | null | t3_1934c3u | /r/LocalLLaMA/comments/1934c3u/slora_available_now_with_openai_v1_completions/ | false | false | 19 | {'enabled': False, 'images': [{'id': 'E7I9hIdyWZ2hG6Loun1LyepmAf9tHBtOHHxZNrySDG8', 'resolutions': [{'height': 54, 'url': 'https://external-preview.redd.it/0jwREb8mXq0qPOM8lMGMGO7FaTOqIQIH7bf3LaKKAUQ.jpg?width=108&crop=smart&auto=webp&s=debc6757514f0a0ea0ab681fd46fa75d953ed1d9', 'width': 108}, {'height': 108, 'url': 'https://external-preview.redd.it/0jwREb8mXq0qPOM8lMGMGO7FaTOqIQIH7bf3LaKKAUQ.jpg?width=216&crop=smart&auto=webp&s=5b33871969387a0fd01363f190ca3c314ec72da7', 'width': 216}, {'height': 160, 'url': 'https://external-preview.redd.it/0jwREb8mXq0qPOM8lMGMGO7FaTOqIQIH7bf3LaKKAUQ.jpg?width=320&crop=smart&auto=webp&s=d31882d8e1705122f8dd5c865f60af66afb965a6', 'width': 320}, {'height': 320, 'url': 'https://external-preview.redd.it/0jwREb8mXq0qPOM8lMGMGO7FaTOqIQIH7bf3LaKKAUQ.jpg?width=640&crop=smart&auto=webp&s=6bc56932665c6b5f8b075c82300db6b949d22308', 'width': 640}, {'height': 480, 'url': 'https://external-preview.redd.it/0jwREb8mXq0qPOM8lMGMGO7FaTOqIQIH7bf3LaKKAUQ.jpg?width=960&crop=smart&auto=webp&s=f479da01a296f3eb1849eb70ad6e9caef1059f56', 'width': 960}, {'height': 540, 'url': 'https://external-preview.redd.it/0jwREb8mXq0qPOM8lMGMGO7FaTOqIQIH7bf3LaKKAUQ.jpg?width=1080&crop=smart&auto=webp&s=b7cc1678f1a859526e3ad706c4be0924bedf3ac9', 'width': 1080}], 'source': {'height': 600, 'url': 'https://external-preview.redd.it/0jwREb8mXq0qPOM8lMGMGO7FaTOqIQIH7bf3LaKKAUQ.jpg?auto=webp&s=695a6b0093f2999328f2a701c7fc664eebee1c15', 'width': 1200}, 'variants': {}}]} | ||
Issues with Q4 K_M of Falcon 180b on Apple Silicon | 1 | Has anyone try this quant of Falcon 180b model on Apple hardware?
Spits out gibberish for me, and I cannot seem to find the problem.
Other models run without any issues. | 2024-01-10T09:13:15 | https://www.reddit.com/r/LocalLLaMA/comments/19346l6/issues_with_q4_k_m_of_falcon_180b_on_apple_silicon/ | Infinite100p | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 19346l6 | false | null | t3_19346l6 | /r/LocalLLaMA/comments/19346l6/issues_with_q4_k_m_of_falcon_180b_on_apple_silicon/ | false | false | self | 1 | null |
New Model: OpenChat 3.5 Update 0106 | 99 | 2024-01-10T08:02:01 | https://twitter.com/openchatdev/status/1744985660870795635 | imonenext | twitter.com | 1970-01-01T00:00:00 | 0 | {} | 193362r | false | {'oembed': {'author_name': 'OpenChat', 'author_url': 'https://twitter.com/openchatdev', 'cache_age': 3153600000, 'height': None, 'html': '<blockquote class="twitter-video"><p lang="en" dir="ltr">🚀Announcing OpenChat-3.5 Update 0106: 𝗪𝗼𝗿𝗹𝗱’𝘀 𝗕𝗲𝘀𝘁 𝗢𝗽𝗲𝗻 𝗦𝗼𝘂𝗿𝗰𝗲 𝟳𝗕 𝗟𝗟𝗠!<br><br>Experience ChatGPT & Grok-level AI locally 💿! <br><br>Surpassing Grok-0 (33B) across all 4 benchmarks and Grok-1 (???B) on average and 3/4 benchmarks 🔥. <br><br>🎯 This update mainly enhanced… <a href="https://t.co/E1XlpRojYp">pic.twitter.com/E1XlpRojYp</a></p>— OpenChat (@openchatdev) <a href="https://twitter.com/openchatdev/status/1744985660870795635?ref_src=twsrc%5Etfw">January 10, 2024</a></blockquote>\n<script async src="https://platform.twitter.com/widgets.js" charset="utf-8"></script>\n\n', 'provider_name': 'Twitter', 'provider_url': 'https://twitter.com', 'type': 'rich', 'url': 'https://twitter.com/openchatdev/status/1744985660870795635', 'version': '1.0', 'width': 350}, 'type': 'twitter.com'} | t3_193362r | /r/LocalLLaMA/comments/193362r/new_model_openchat_35_update_0106/ | false | false | 99 | {'enabled': False, 'images': [{'id': '1DmMCM7cxnDVP_kNLTGy3khmzwKR1Mnvby36DIMiQU4', 'resolutions': [{'height': 54, 'url': 'https://external-preview.redd.it/Z-MT5qMk5_PAo84Oosqf4Lv0cj-iV23xqAPliBpvTl4.jpg?width=108&crop=smart&auto=webp&s=91e2c20e4983532a5771736065ba96b20d1479be', 'width': 108}], 'source': {'height': 70, 'url': 'https://external-preview.redd.it/Z-MT5qMk5_PAo84Oosqf4Lv0cj-iV23xqAPliBpvTl4.jpg?auto=webp&s=547299d71797ffeab94ebbafaf213bb0295c35ed', 'width': 140}, 'variants': {}}]} | ||
benchmarks with different temps, seeds? | 4 | Temperature can influence greatly the outcome of your response, seed also.
I've seen LLMs solve different problems with lower temps, some other more creative tasks with higher temps. I believe we need a way to adjust temperature to the prompt somehow. And also, why not generate 100 responses, make it analyse all 100 responses, and pick 1. Sometimes they can be genius about some things, and it's hard to 'link' it but once the link is made you see it and easy to recognize it.
When you are asked a tough question. You explore things in your mind, you come up with several ideas, some are bad, some good, you synthetise, you improve on your idea, etc.
I've seen stupid LLMs (13B) that were pretty ingenious sometimes, if the parameters were right. And we fall into this trap, thinking that if a LLM gives 10 responses bad, it is uncapable of coming up with a genius response or it is trash.
I think it's important that when we run benchmarks to have these parameters taken into account. And why not generate for 10k seeds.
\---
Here's what I would like:
\- benchmarks on different temps ranges (0.7 1 1.25) - I will let the experts decide
\- benchmarks that run it on 100 sample seeds, and have a way to select its best response somehow.
​
Because again, playing locally on 2x 3090 with some exl2 models, is very fun, and they are very intelligent and creative and they can be finetuned, they can regenerate responses if you don't like, or if you want to explore other paths.
thoughts?
​ | 2024-01-10T07:59:15 | https://www.reddit.com/r/LocalLLaMA/comments/19334dm/benchmarks_with_different_temps_seeds/ | shaman-warrior | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 19334dm | false | null | t3_19334dm | /r/LocalLLaMA/comments/19334dm/benchmarks_with_different_temps_seeds/ | false | false | self | 4 | null |
OpenChat 3.5 Update 0106 | 1 | 🚀Announcing OpenChat-3.5 Update 0106: 𝗪𝗼𝗿𝗹𝗱’𝘀 𝗕𝗲𝘀𝘁 𝗢𝗽𝗲𝗻 𝗦𝗼𝘂𝗿𝗰𝗲 𝟳𝗕 𝗟𝗟𝗠!
Experience ChatGPT & Grok-level AI locally 💿!
Surpassing Grok-0 (33B) across all 4 benchmarks and Grok-1 (???B) on average and 3/4 benchmarks 🔥.
https://preview.redd.it/bmbzts9ogkbc1.png?width=1920&format=png&auto=webp&s=de945db73660e455615d526e35d2afc875f8751e
🎯 This update mainly enhanced training methodology, in-context learning & coding skills, outperforming the last 1210 release on 7 out of 8 benchmarks!
​
🌐 The model is available on our live demo, HuggingFace, and GitHub:
​
HuggingFace: [https://huggingface.co/openchat/openchat-3.5-0106](https://huggingface.co/openchat/openchat-3.5-0106)
Live Demo: [https://openchat.team](https://openchat.team)
GitHub: [https://github.com/imoneoi/openchat](https://github.com/imoneoi/openchat)
​
🛠️ To deploy it yourself, visit our GitHub ([https://github.com/imoneoi/openchat](https://github.com/imoneoi/openchat)) for full instructions to serve OpenChat models with an accelerated vLLM backend, API key authentication, and more! | 2024-01-10T07:41:05 | https://www.reddit.com/r/LocalLLaMA/comments/1932v5c/openchat_35_update_0106/ | imonenext | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 1932v5c | false | null | t3_1932v5c | /r/LocalLLaMA/comments/1932v5c/openchat_35_update_0106/ | false | false | 1 | {'enabled': False, 'images': [{'id': 'RhdvJZieIzJqk3S4MQFhDhduFk-urTLEG-k6Yu5Z5D0', 'resolutions': [{'height': 58, 'url': 'https://external-preview.redd.it/thqTyNpDS0w8kJiIOitvGsrgE-eu1lU0H9T_wiuQsZw.jpg?width=108&crop=smart&auto=webp&s=59ee82ef3de568b6850e3c0f92085881081018e4', 'width': 108}, {'height': 116, 'url': 'https://external-preview.redd.it/thqTyNpDS0w8kJiIOitvGsrgE-eu1lU0H9T_wiuQsZw.jpg?width=216&crop=smart&auto=webp&s=45b68f858ea7aee06b860203a8c502c177f07fcd', 'width': 216}, {'height': 172, 'url': 'https://external-preview.redd.it/thqTyNpDS0w8kJiIOitvGsrgE-eu1lU0H9T_wiuQsZw.jpg?width=320&crop=smart&auto=webp&s=b13a0aaca81eec85025a6a22de95e1b710bcb3d3', 'width': 320}, {'height': 345, 'url': 'https://external-preview.redd.it/thqTyNpDS0w8kJiIOitvGsrgE-eu1lU0H9T_wiuQsZw.jpg?width=640&crop=smart&auto=webp&s=0732bf40f1dcd981ddff1f90a5d08cf571e66a28', 'width': 640}, {'height': 518, 'url': 'https://external-preview.redd.it/thqTyNpDS0w8kJiIOitvGsrgE-eu1lU0H9T_wiuQsZw.jpg?width=960&crop=smart&auto=webp&s=da5f6cd9c7b7b211b9cd93b8db6aa3d715cc6367', 'width': 960}, {'height': 583, 'url': 'https://external-preview.redd.it/thqTyNpDS0w8kJiIOitvGsrgE-eu1lU0H9T_wiuQsZw.jpg?width=1080&crop=smart&auto=webp&s=e0b69f2cea31bfea0f431cba909b5acfe20529b6', 'width': 1080}], 'source': {'height': 648, 'url': 'https://external-preview.redd.it/thqTyNpDS0w8kJiIOitvGsrgE-eu1lU0H9T_wiuQsZw.jpg?auto=webp&s=a41b8b3a04b35a6a8ee2acb8fee736bb697b2038', 'width': 1200}, 'variants': {}}]} | |
Can anyone explain the Mamba space state model to a total noob? | 1 | [removed] | 2024-01-10T07:38:00 | https://www.reddit.com/r/LocalLLaMA/comments/1932ti6/can_anyone_explain_the_mamba_space_state_model_to/ | Fun_Cat3671 | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 1932ti6 | false | null | t3_1932ti6 | /r/LocalLLaMA/comments/1932ti6/can_anyone_explain_the_mamba_space_state_model_to/ | false | false | self | 1 | null |
Help | 1 | Hi
I have subscribed with chingu.fun to see how it work. But found really bad, nothing new, run on openai.
But there is no way to cancel my subscription. Thel help line has only email.
How i will cancel this subscription? Please help. | 2024-01-10T07:33:07 | https://www.reddit.com/r/LocalLLaMA/comments/1932qz3/help/ | Beanbag887 | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 1932qz3 | false | null | t3_1932qz3 | /r/LocalLLaMA/comments/1932qz3/help/ | false | false | self | 1 | null |
Hardware Specs for Quick 7b or 13b Model Responses | 6 | What hardware specifications are recommended to efficiently run the 7b or 13b models, ensuring responses within 2 to 3 seconds, and what components contribute to achieving this performance level? | 2024-01-10T07:12:15 | https://www.reddit.com/r/LocalLLaMA/comments/1932fok/hardware_specs_for_quick_7b_or_13b_model_responses/ | East-Bug6675 | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 1932fok | false | null | t3_1932fok | /r/LocalLLaMA/comments/1932fok/hardware_specs_for_quick_7b_or_13b_model_responses/ | false | false | self | 6 | null |
The problem with LLMs and NLP ? | 1 | [removed] | 2024-01-10T06:18:52 | https://www.reddit.com/r/LocalLLaMA/comments/1931krw/the_problem_with_llms_and_nlp/ | SnooBeans7516 | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 1931krw | false | null | t3_1931krw | /r/LocalLLaMA/comments/1931krw/the_problem_with_llms_and_nlp/ | false | false | self | 1 | {'enabled': False, 'images': [{'id': 'PLI-AOVL2mHF2MAr5_JA_Tg-67e1I3ju1PpdhkBEDqc', 'resolutions': [{'height': 50, 'url': 'https://external-preview.redd.it/XPYC2WDjhL_Gbu-YRU-8YiO9m5Tp5ONrHomhd51Nplk.jpg?width=108&crop=smart&auto=webp&s=51a46ec82439bf2f0129657e14f17db2d89c8032', 'width': 108}, {'height': 100, 'url': 'https://external-preview.redd.it/XPYC2WDjhL_Gbu-YRU-8YiO9m5Tp5ONrHomhd51Nplk.jpg?width=216&crop=smart&auto=webp&s=899650b74b00453db02e6f8fab7b2cbed9372d8b', 'width': 216}, {'height': 148, 'url': 'https://external-preview.redd.it/XPYC2WDjhL_Gbu-YRU-8YiO9m5Tp5ONrHomhd51Nplk.jpg?width=320&crop=smart&auto=webp&s=ee7b077d05b5d5faa61b60591a11d461e181df97', 'width': 320}, {'height': 296, 'url': 'https://external-preview.redd.it/XPYC2WDjhL_Gbu-YRU-8YiO9m5Tp5ONrHomhd51Nplk.jpg?width=640&crop=smart&auto=webp&s=474aa6784fb8b7520cda5b95e42bed645e9c9917', 'width': 640}, {'height': 444, 'url': 'https://external-preview.redd.it/XPYC2WDjhL_Gbu-YRU-8YiO9m5Tp5ONrHomhd51Nplk.jpg?width=960&crop=smart&auto=webp&s=1d38b984256f364dec78a9c73c53e79fefed2582', 'width': 960}, {'height': 500, 'url': 'https://external-preview.redd.it/XPYC2WDjhL_Gbu-YRU-8YiO9m5Tp5ONrHomhd51Nplk.jpg?width=1080&crop=smart&auto=webp&s=60225b614fdc8957c93287839c0891bae522f1d6', 'width': 1080}], 'source': {'height': 556, 'url': 'https://external-preview.redd.it/XPYC2WDjhL_Gbu-YRU-8YiO9m5Tp5ONrHomhd51Nplk.jpg?auto=webp&s=38f753fd74d3abcc94282e77613f155c53874bad', 'width': 1200}, 'variants': {}}]} |
Are there any good 2b models for basic chatbot ? | 1 | [removed] | 2024-01-10T06:00:20 | https://www.reddit.com/r/LocalLLaMA/comments/1931986/are_there_any_good_2b_models_for_basic_chatbot/ | AnabelBain | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 1931986 | false | null | t3_1931986 | /r/LocalLLaMA/comments/1931986/are_there_any_good_2b_models_for_basic_chatbot/ | false | false | self | 1 | null |
Can anyone tell me tok/sec for 4x P40's in an old PCIe3 server? | 3 | The title pretty much says it all.
I'm running the 103B models at Q6\_K in an old dell server using \~95gb of DDR3 ram and a couple of old Xeons (no AVX2). The original plan was to offload some of the layers onto a couple of 3090's but with the PCIe3 bus, I wondered if I would be better off with 4x P40's and offloading all the layers.
The other concern I have is that llama.cpp might stop supporting old Pascal but I'm not sure anyone will know if that is likely.
In case anyone is curious, I'm getting a blistering 0.2 tok/sec using CPU right now so that's the number to beat. It is very much a "request a few thousand tokens overnight" kind of batch setup right now.
Thanks in advance. | 2024-01-10T05:57:17 | https://www.reddit.com/r/LocalLLaMA/comments/193177o/can_anyone_tell_me_toksec_for_4x_p40s_in_an_old/ | maz_net_au | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 193177o | false | null | t3_193177o | /r/LocalLLaMA/comments/193177o/can_anyone_tell_me_toksec_for_4x_p40s_in_an_old/ | false | false | self | 3 | null |
Deepseek Coder MoE = Best python LLM? | 4 | I am looking for collaborators to build an MoE model using Deepseek Coder. The goal is to build the best python LLM. Deepseek Coder 7b already beats GPT-3.5, and with MoE, we can close the gap with GPT-4. + It also has very permissive license. | 2024-01-10T05:28:08 | https://www.reddit.com/r/LocalLLaMA/comments/1930ohn/deepseek_coder_moe_best_python_llm/ | anubrag | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 1930ohn | false | null | t3_1930ohn | /r/LocalLLaMA/comments/1930ohn/deepseek_coder_moe_best_python_llm/ | false | false | self | 4 | null |
Deepmoney: A High-End LLM in finance based on massive research data | 61 | Happy 2024! Guys! I've trained a lot of models over the past year, but a lot of the work went into this model. The data and computing power used in this model are of the highest quality I have ever used since I came into contact with large models.
Base model: [TriadParty/deepmoney-34b-200k-base · Hugging Face](https://huggingface.co/TriadParty/deepmoney-34b-200k-base)
Chat model( In fact, it was designed as an agent for me to build quantitative analysis. ): [TriadParty/deepmoney-34b-200k-chat-evaluator · Hugging Face](https://huggingface.co/TriadParty/deepmoney-34b-200k-chat-evaluator)
There are many technical details used, you can go to the link to see the details. But if you have any questions, you can ask me here anytime! | 2024-01-10T05:04:55 | https://www.reddit.com/r/LocalLLaMA/comments/19308rg/deepmoney_a_highend_llm_in_finance_based_on/ | Fun_Water2230 | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 19308rg | false | null | t3_19308rg | /r/LocalLLaMA/comments/19308rg/deepmoney_a_highend_llm_in_finance_based_on/ | false | false | self | 61 | {'enabled': False, 'images': [{'id': 'CPcMpE8BjLkSLb_X3XZgJtXLdQIfyd9QAYIcT7xw0Qk', 'resolutions': [{'height': 58, 'url': 'https://external-preview.redd.it/oPYMJED6rf3rSTqCQfAmNhe5YAjTQVufIQvCGRW81QE.jpg?width=108&crop=smart&auto=webp&s=6f4adda118f3e50881e930f27e45d23b2615eed7', 'width': 108}, {'height': 116, 'url': 'https://external-preview.redd.it/oPYMJED6rf3rSTqCQfAmNhe5YAjTQVufIQvCGRW81QE.jpg?width=216&crop=smart&auto=webp&s=f003323c5e98416ca3851c74421b064672f23240', 'width': 216}, {'height': 172, 'url': 'https://external-preview.redd.it/oPYMJED6rf3rSTqCQfAmNhe5YAjTQVufIQvCGRW81QE.jpg?width=320&crop=smart&auto=webp&s=0f6b38e5a41c3d5bb04568e77ac828cbdb0c8038', 'width': 320}, {'height': 345, 'url': 'https://external-preview.redd.it/oPYMJED6rf3rSTqCQfAmNhe5YAjTQVufIQvCGRW81QE.jpg?width=640&crop=smart&auto=webp&s=345ea124ad83f3c1536c25efc1c58da2244ce9f4', 'width': 640}, {'height': 518, 'url': 'https://external-preview.redd.it/oPYMJED6rf3rSTqCQfAmNhe5YAjTQVufIQvCGRW81QE.jpg?width=960&crop=smart&auto=webp&s=4e58151762c0e71004deb25c0cb703b3e08d92b2', 'width': 960}, {'height': 583, 'url': 'https://external-preview.redd.it/oPYMJED6rf3rSTqCQfAmNhe5YAjTQVufIQvCGRW81QE.jpg?width=1080&crop=smart&auto=webp&s=ca0c42c815818216b6d90d616d7e6cdb1141204d', 'width': 1080}], 'source': {'height': 648, 'url': 'https://external-preview.redd.it/oPYMJED6rf3rSTqCQfAmNhe5YAjTQVufIQvCGRW81QE.jpg?auto=webp&s=23bbb1b64e3bb79df69b1489a8aa02415ff2fd83', 'width': 1200}, 'variants': {}}]} |
Mixtral token and experts explanataion | 1 | Hi
I was reading how the Mixtral 8x7b works and I am trying to achieve some basic understanding in terms of how the expert models are used. So according to the recently published paper, Mixtral was trained to best work with 2 model experts per token selected by the router (out of 8 in total). I would like to understand this part better, so how exactly 2 experts per token work? Does it mean for each token in the prompt there are assigned two experts to read the token and then produce a new one? Each token can have different experts or is it the same pair of experts for whole prompt that both reads the prompt and then generates new output? Or maybe each token has separate experts no matter whether it is reading or writing new tokens? How does this work in practice? Appreciate simple explanation for this :) | 2024-01-10T04:36:35 | https://www.reddit.com/r/LocalLLaMA/comments/192zoyn/mixtral_token_and_experts_explanataion/ | investopim | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 192zoyn | false | null | t3_192zoyn | /r/LocalLLaMA/comments/192zoyn/mixtral_token_and_experts_explanataion/ | false | false | self | 1 | null |
I know nobody cares when 'AI's say the darndest things', but this got me. | 15 | I asked a model to tell me a bar joke. The jokes they tell are usually bad, don't make sense, etc, but this cracked me up for some reason.
​
https://preview.redd.it/mdq53c0e9jbc1.png?width=806&format=png&auto=webp&s=666e4f515ea4f7544fe698cc9c9c56e82d446ad2 | 2024-01-10T03:36:14 | https://www.reddit.com/r/LocalLLaMA/comments/192yk1r/i_know_nobody_cares_when_ais_say_the_darndest/ | multiverse_fan | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 192yk1r | false | null | t3_192yk1r | /r/LocalLLaMA/comments/192yk1r/i_know_nobody_cares_when_ais_say_the_darndest/ | false | false | 15 | null | |
Mixtral 34bx2 - is it REALLY slow for anyone else? | 17 | I'm using an M2 Mac Ultra Mac Studio and using Oobabooga for inference, and tonight I loaded up that model Mixtral 34bx2. I thought that perhaps with the smaller size, it might run a little more quickly than 70b model... but holy crap was I wrong.
* Goliath 120b q8 with 6144 Context Window: Responds in 150-170 seconds.
* Mixtral 34bx2 q8 with 6144 Context Window: Responds in 260-280 seconds.
* Mixtral 34bx2 q8 with 16k Context Window: Responds in 670 seconds (that's 11 minutes...)
I'm really confused how my slowest model, Goliath 120b, is FASTER than the Mixtral 34bx2.
Has anyone else noticed similar? | 2024-01-10T03:30:31 | https://www.reddit.com/r/LocalLLaMA/comments/192yg5q/mixtral_34bx2_is_it_really_slow_for_anyone_else/ | SomeOddCodeGuy | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 192yg5q | false | null | t3_192yg5q | /r/LocalLLaMA/comments/192yg5q/mixtral_34bx2_is_it_really_slow_for_anyone_else/ | false | false | self | 17 | null |
We created an app for all your AI services. | 1 | [removed] | 2024-01-10T03:26:57 | https://www.reddit.com/r/LocalLLaMA/comments/192ydi1/we_created_an_app_for_all_your_ai_services/ | Any_Salamander_2058 | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 192ydi1 | false | null | t3_192ydi1 | /r/LocalLLaMA/comments/192ydi1/we_created_an_app_for_all_your_ai_services/ | false | false | self | 1 | {'enabled': False, 'images': [{'id': 'FBWwmAZRNYY0ZQHu3mGXZ9p6ewu_0yO2a9rnxifeYqs', 'resolutions': [{'height': 56, 'url': 'https://external-preview.redd.it/V0FBs4ANBGpyFs_PZ9-yzcRHWxeq6OF-hv624Ti8Fs0.jpg?width=108&crop=smart&auto=webp&s=76adf9aa07171352e3727b8d8bde812b5eb0f7ab', 'width': 108}, {'height': 113, 'url': 'https://external-preview.redd.it/V0FBs4ANBGpyFs_PZ9-yzcRHWxeq6OF-hv624Ti8Fs0.jpg?width=216&crop=smart&auto=webp&s=f180f4caf096ee042b700d70cabf941788671fda', 'width': 216}, {'height': 168, 'url': 'https://external-preview.redd.it/V0FBs4ANBGpyFs_PZ9-yzcRHWxeq6OF-hv624Ti8Fs0.jpg?width=320&crop=smart&auto=webp&s=204c440e43a68783d15356bdc60ce239349a1d9f', 'width': 320}, {'height': 336, 'url': 'https://external-preview.redd.it/V0FBs4ANBGpyFs_PZ9-yzcRHWxeq6OF-hv624Ti8Fs0.jpg?width=640&crop=smart&auto=webp&s=2689ba6aac51113255e3e6e8acb33870740831e1', 'width': 640}, {'height': 504, 'url': 'https://external-preview.redd.it/V0FBs4ANBGpyFs_PZ9-yzcRHWxeq6OF-hv624Ti8Fs0.jpg?width=960&crop=smart&auto=webp&s=64ad438b94a08899677dfb6f2de708f8a7a6a81b', 'width': 960}, {'height': 567, 'url': 'https://external-preview.redd.it/V0FBs4ANBGpyFs_PZ9-yzcRHWxeq6OF-hv624Ti8Fs0.jpg?width=1080&crop=smart&auto=webp&s=0108876ef023c871780ce4ec2584fa4cb25c499c', 'width': 1080}], 'source': {'height': 630, 'url': 'https://external-preview.redd.it/V0FBs4ANBGpyFs_PZ9-yzcRHWxeq6OF-hv624Ti8Fs0.jpg?auto=webp&s=9035877a533fae2ee246bfc56769d60e805f7b74', 'width': 1200}, 'variants': {}}]} |
Running Ferret / Visual vicuna. | 6 | Anyone around here ever succesfully run ml-ferret from Apple?
After an entire day of sweat and learning, I finally got https://github.com/apple/ml-ferret to run its demo.
Unfortunately, when I try one of the examples in the demo web page, I get this error:
```
Caught Unknown Error CUDA out of memory. Tried to allocate 38.00 MiB. GPU 0 has a total capacty of 11.76 GiB of which 28.31 MiB is free. Including non-PyTorch memory, this process has 10.13 GiB memory in use. Of the allocated memory 9.94 GiB is allocated by PyTorch, and 52.19 MiB is reserved by PyTorch but unallocated. If reserved but unallocated memory is large try setting max_split_size_mb to avoid fragmentation. See documentation for Memory Management and PYTORCH_CUDA_ALLOC_CONF
```
I have a 3060 with about 10GB of usable VRAM, I thought it'd be enough.
My guess is it's running the model in 32bits, so trying to take up too much space.
I'd like to try running it in 8bit or 4bit, but i have no idea how.
The model server has a --load-4bit option, but that doesn't work.
```
2024-01-10 03:23:54 | ERROR | stderr | TypeError: BitsAndBytesConfig.__init__() got an unexpected keyword argument 'load_in_4bit'
```
My best guess would be, I need to do **something** to the model before its able to run in 4bit, quantize it, but I have no idea how, and I wasn't able to find any ferret-specific documentation around.
Anyone here has been able to run ferret?
If so, any help would be immensely appreciated.
Thanks!!
Cheers. | 2024-01-10T02:25:39 | https://www.reddit.com/r/LocalLLaMA/comments/192x4at/running_ferret_visual_vicuna/ | arthurwolf | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 192x4at | false | null | t3_192x4at | /r/LocalLLaMA/comments/192x4at/running_ferret_visual_vicuna/ | false | false | self | 6 | {'enabled': False, 'images': [{'id': 'ArjTuUtmv1iPW2E5rmJ1k-f_ohNW5mipCbTCUo7-f5s', 'resolutions': [{'height': 54, 'url': 'https://external-preview.redd.it/ugWvhYUKaESGVuPQmgce_Yi9CDBrjd0EDuKo2Ai4GrQ.jpg?width=108&crop=smart&auto=webp&s=f96cb2570c0a733346b215a8fd4758804c1fc965', 'width': 108}, {'height': 108, 'url': 'https://external-preview.redd.it/ugWvhYUKaESGVuPQmgce_Yi9CDBrjd0EDuKo2Ai4GrQ.jpg?width=216&crop=smart&auto=webp&s=f52d8debc429c534cf95de3cd3bf2ff5e06fc10f', 'width': 216}, {'height': 160, 'url': 'https://external-preview.redd.it/ugWvhYUKaESGVuPQmgce_Yi9CDBrjd0EDuKo2Ai4GrQ.jpg?width=320&crop=smart&auto=webp&s=98808b5272b90c770f9b4fb81722368ff39f709a', 'width': 320}, {'height': 320, 'url': 'https://external-preview.redd.it/ugWvhYUKaESGVuPQmgce_Yi9CDBrjd0EDuKo2Ai4GrQ.jpg?width=640&crop=smart&auto=webp&s=f2cc8c55ca16fb3540bb9093ccc0390941e15a63', 'width': 640}, {'height': 480, 'url': 'https://external-preview.redd.it/ugWvhYUKaESGVuPQmgce_Yi9CDBrjd0EDuKo2Ai4GrQ.jpg?width=960&crop=smart&auto=webp&s=da8b04b497aedd3471c38eb3da9de10ded36f822', 'width': 960}, {'height': 540, 'url': 'https://external-preview.redd.it/ugWvhYUKaESGVuPQmgce_Yi9CDBrjd0EDuKo2Ai4GrQ.jpg?width=1080&crop=smart&auto=webp&s=98d41fc1ca721aea54c897a6e5d8fe122693e4cb', 'width': 1080}], 'source': {'height': 600, 'url': 'https://external-preview.redd.it/ugWvhYUKaESGVuPQmgce_Yi9CDBrjd0EDuKo2Ai4GrQ.jpg?auto=webp&s=d6a9d3a712a699658c8b81c59c9c9abfcea29448', 'width': 1200}, 'variants': {}}]} |
Anything better than goat-70b-storytelling in the <50GB range? | 10 | This has been my go to model for collaborative story writing and general questions and answers. Curious if anyone has any recommendations for something better.
I use kcpp, if that matters. | 2024-01-10T02:25:33 | https://www.reddit.com/r/LocalLLaMA/comments/192x48l/anything_better_than_goat70bstorytelling_in_the/ | wh33t | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 192x48l | false | null | t3_192x48l | /r/LocalLLaMA/comments/192x48l/anything_better_than_goat70bstorytelling_in_the/ | false | false | self | 10 | null |
Phixtral: Mixture of Experts Models with Phi | 94 | 2024-01-10T02:06:36 | https://x.com/maximelabonne/status/1744867841436700850?s=20 | ninjasaid13 | x.com | 1970-01-01T00:00:00 | 0 | {} | 192wq14 | false | null | t3_192wq14 | /r/LocalLLaMA/comments/192wq14/phixtral_mixture_of_experts_models_with_phi/ | false | false | 94 | {'enabled': False, 'images': [{'id': 'co46qWWkkHDFTiNRlRJQ6BEvrt3pbOPKg1JsBoNQ44E', 'resolutions': [{'height': 60, 'url': 'https://external-preview.redd.it/xZ_vWEh7F5BpY4dXipMqC2v6mO5jYF--fL0tDes74BE.jpg?width=108&crop=smart&auto=webp&s=cb5ec86938d85a732fc4350615dd68fcd314ea85', 'width': 108}, {'height': 121, 'url': 'https://external-preview.redd.it/xZ_vWEh7F5BpY4dXipMqC2v6mO5jYF--fL0tDes74BE.jpg?width=216&crop=smart&auto=webp&s=a5fab0ea3c18f42d9ff8581311f316641aa69ffd', 'width': 216}, {'height': 180, 'url': 'https://external-preview.redd.it/xZ_vWEh7F5BpY4dXipMqC2v6mO5jYF--fL0tDes74BE.jpg?width=320&crop=smart&auto=webp&s=a678972803c24f0b60648d6aea1df1d7015a5d86', 'width': 320}, {'height': 360, 'url': 'https://external-preview.redd.it/xZ_vWEh7F5BpY4dXipMqC2v6mO5jYF--fL0tDes74BE.jpg?width=640&crop=smart&auto=webp&s=ded22bcbae301fa7a0e397885f8ddf5f1b26f379', 'width': 640}, {'height': 540, 'url': 'https://external-preview.redd.it/xZ_vWEh7F5BpY4dXipMqC2v6mO5jYF--fL0tDes74BE.jpg?width=960&crop=smart&auto=webp&s=ea62d4def900602f0c09cfb070fc1840635028eb', 'width': 960}, {'height': 607, 'url': 'https://external-preview.redd.it/xZ_vWEh7F5BpY4dXipMqC2v6mO5jYF--fL0tDes74BE.jpg?width=1080&crop=smart&auto=webp&s=3cdb4ba7afc7ff1c187000dbcc0553ce75daa669', 'width': 1080}], 'source': {'height': 1080, 'url': 'https://external-preview.redd.it/xZ_vWEh7F5BpY4dXipMqC2v6mO5jYF--fL0tDes74BE.jpg?auto=webp&s=e7541ac07fba837b70d04020f3c5a7f6e27f6668', 'width': 1920}, 'variants': {}}]} | ||
Efficient Batching | 11 | I came up with a novel way to do efficient batching.
I basically permutate a list of strings
identify their lengths
and iterate down through the list (removing selected for records from available records) until block\_size is filled
and if there are no more records that can fill it, pop the first record out adding it back into the available ones to choose from and pick the next until the sequence is filled.
The result is a single batch up to block\_size in length, novel generation, guaranteed to be unique and not repeating within that batch as well as effectively give consideration to all records equally
[https://gist.github.com/thistleknot/97617f91538ad075b9a44437f88e8680](https://gist.github.com/thistleknot/97617f91538ad075b9a44437f88e8680) | 2024-01-10T02:05:53 | https://www.reddit.com/r/LocalLLaMA/comments/192wphf/efficient_batching/ | Thistleknot | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 192wphf | false | null | t3_192wphf | /r/LocalLLaMA/comments/192wphf/efficient_batching/ | false | false | self | 11 | {'enabled': False, 'images': [{'id': 'OAXSl8SY6T3JK9MGQyKxkoYbqZ71HQRYXLeB8CV0NXg', 'resolutions': [{'height': 54, 'url': 'https://external-preview.redd.it/DaucjXMGsNHM-CtmdilC9-Be6MC8V2z4ykjVCgOkTFc.jpg?width=108&crop=smart&auto=webp&s=9bcab7b79864ff27bf48116cb335a6f825bfb124', 'width': 108}, {'height': 108, 'url': 'https://external-preview.redd.it/DaucjXMGsNHM-CtmdilC9-Be6MC8V2z4ykjVCgOkTFc.jpg?width=216&crop=smart&auto=webp&s=e4e925345605c644eebe8abd69916915fc4fbcf7', 'width': 216}, {'height': 160, 'url': 'https://external-preview.redd.it/DaucjXMGsNHM-CtmdilC9-Be6MC8V2z4ykjVCgOkTFc.jpg?width=320&crop=smart&auto=webp&s=614b06d5b40c890a59e355191a6e2d75cdf50789', 'width': 320}, {'height': 320, 'url': 'https://external-preview.redd.it/DaucjXMGsNHM-CtmdilC9-Be6MC8V2z4ykjVCgOkTFc.jpg?width=640&crop=smart&auto=webp&s=62ca4cb88917f17e7200a6f1c665b5d959713745', 'width': 640}, {'height': 480, 'url': 'https://external-preview.redd.it/DaucjXMGsNHM-CtmdilC9-Be6MC8V2z4ykjVCgOkTFc.jpg?width=960&crop=smart&auto=webp&s=c5f4a30974a8e6bad0d617a79935bc70c954e3e8', 'width': 960}, {'height': 540, 'url': 'https://external-preview.redd.it/DaucjXMGsNHM-CtmdilC9-Be6MC8V2z4ykjVCgOkTFc.jpg?width=1080&crop=smart&auto=webp&s=476793be11eaac4604b6b0c938b45c7c3b52d450', 'width': 1080}], 'source': {'height': 640, 'url': 'https://external-preview.redd.it/DaucjXMGsNHM-CtmdilC9-Be6MC8V2z4ykjVCgOkTFc.jpg?auto=webp&s=9ae035fbdcd6bb503ab0b4a605b8db6de46647ee', 'width': 1280}, 'variants': {}}]} |
I am extremely confused. Follow-up post to yesterday’s “I fried my desktop using TextGen WebUI, come laugh at me” post | 28 | First off, I saw a couple of people assuming it was a joke. I thoroughly wish it was. Some guys blow money on cars, others on hookers or meth… my rig is my child. I built this Frankenstein with my bare hands and my life revolves around it. The last 24 hours have been quite stressful. So, while I dressed it up this disaster to sound funny for a reddit post, let me assure you, I wasn’t laughing.
Now, I honestly have no idea what the heck just happened, but I vaguely understand what I did wrong. What I don’t understand is the -specifics- or how it got fixed. So, if you’re good at diagnosing obscure PC problems… any advice would be greatly appreciated.
Here’s the timeline of events.
Original setup, as of two weeks ago:
AORUS 570 motherboard
Ryzen 9 3900X
RTX 2060, 12GB VRAM
64GB RAM, DDR4 (4x 16)
Which I then upgraded to—
RTX 4060 TI 16GB VRAM
That felt beautiful but I wanted more. So a week later I switched the RAM to DDR4, 128 GB RAM. (4x 32)
And for a whole three days everything was perfect. Happy camper, running Goliath 120B 5K_M_GGUF and basically whatever else I wanted to fool around with.
—
Now, skipping to the dumb user event that caused trouble. I wanted to close out of TextGeneration WebUI and was not happy with how fast it was exiting the program, so, mid prompt I rebooted my PC. From that point on I got blue screens. Could not log back into my computer at all. I eventually took 3 RAM sticks out and got it working in safe mode, and, decided to update my drivers praying that was the issue. That made it worse somehow.
Mind you: I also tested using the previous 64GB RAM sticks, which thankfully had not been sold yet, so I could get some insights on the issue here. So, at that point, it went from being able to run 128GB RAM, no issues, to running 32GB RAM in safe mode, to blue-screening everything.
My only option at that point was to do a full Windows reset. I did so.
Now here’s where it gets unbelievably bizarre.
I rebooted the PC, freshly reset on one RAM stick. It worked. I loaded it with four. Blue screens, the problem came back.
I then took three out, loaded it again. Booted fine. Loaded up a second RAM stick. Still works. Loaded the third… still fine.
Loaded the fourth and…. Somehow I now have a working PC at 128GB RAM again. To test it I’ve run Stockfish 16 on all 24 cores for thirty minutes straight and if that doesn’t break a machine then it’s probably working.
So… my questions are this:
What was the fatal issue, and how did it magically fix itself, albeit after a full Windows reset?
I’m so confused. | 2024-01-10T02:05:49 | https://www.reddit.com/r/LocalLLaMA/comments/192wpfo/i_am_extremely_confused_followup_post_to/ | ThrowawayEmail000 | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 192wpfo | false | null | t3_192wpfo | /r/LocalLLaMA/comments/192wpfo/i_am_extremely_confused_followup_post_to/ | false | false | self | 28 | null |
Is there a way to inspect the contents of LLM data files? | 2 | Apologies for the noob question. I'm very new to LLMs.
I had a lively argument with a bunch of people complaining LLMs are stealing their work. One of their claims is that LLM "store knowledge" and provided compelling examples as "proof".
To settle this argument, I spun out ollama on docker and ran both mistral and mixtral models. I also downloaded the 7Gb mistral tar file.
The idea is to ask some general knowledge questions to the model, compare to what search engines return, then inspect the data downloaded by ollama, or the files from the tar file, to see if such text is stored verbatim.
This last step is where i'm stuck. The tar file contains like 30 450Mb binary files in a format I can't find any documentation about.
Is there a tool that can read the binary format so I can search for such "stored knowledge"? | 2024-01-10T01:20:07 | https://www.reddit.com/r/LocalLLaMA/comments/192vqom/is_there_a_way_to_inspect_the_contents_of_llm/ | machaqueso | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 192vqom | false | null | t3_192vqom | /r/LocalLLaMA/comments/192vqom/is_there_a_way_to_inspect_the_contents_of_llm/ | false | false | self | 2 | null |
Exploring the best approach for AI Copilot for Doctors - Which specific Models to use and suggestion for Hardware | 1 | I'm working on a project focused on developing an AI Copilot for doctors. The idea is to create a digital assistant that can automatically generate clinical notes from conversations between patients and doctors.
In my initial exploration, I've been contemplating the use of GPT-3.5-turbo in a different project, I'm considering its potential application in this healthcare context as well. However, due to the limited budget and high token usage I'm exploring opensource solutions now.
Currently, I'm working a server with RTX 3090, which might be relatively slow for this task, I've also been looking into LM Studio for its API, but I'm in the early stages of figuring out the most efficient approach.
I'm reaching out to the community for advice on the best models or hardware configurations for handling this type of project efficiently. I've thought about exploring server rentals,but I prefer short, flexible sessions during the evenings.
If anyone has experience or suggestions regarding alternative methods or models that could run efficiently on my existing hardware, I would greatly appreciate your insights.
Also, I'm also looking for a co-founder for the project as well, If anyone interested, DM me.
​ | 2024-01-10T01:02:20 | https://www.reddit.com/r/LocalLLaMA/comments/192vcm2/exploring_the_best_approach_for_ai_copilot_for/ | thereisnospooongeek | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 192vcm2 | false | null | t3_192vcm2 | /r/LocalLLaMA/comments/192vcm2/exploring_the_best_approach_for_ai_copilot_for/ | false | false | self | 1 | null |
Extremely slow speeds | 1 | I am having trouble getting text generation to even get to 1 token/s. I am using a 4070 TI and have confirmed that it is using the GPU and has CUDA. I'm running it on Windows and using oobabooga, trying to run 7-13b models with 4-8bit. Any thoughts on what might be going wrong and what I could do to fix it? | 2024-01-10T00:51:56 | https://www.reddit.com/r/LocalLLaMA/comments/192v3zb/extremely_slow_speeds/ | ThisSiteIsShitMan | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 192v3zb | false | null | t3_192v3zb | /r/LocalLLaMA/comments/192v3zb/extremely_slow_speeds/ | false | false | self | 1 | null |
ISO: Smallest most capable Summarization model | 15 | As the title says, I'm looking for the smallest most capable Summarization model.
When I'm looking for models with certain capabilities I usually search huggingface, but on huggingface you don't often get real opinions from real people about model capabilities as on reddit.
I'll take any leaderboards with a grain of salt because while some models are feed to meet specific metrics, it still meets said metric.
Seeing as how Summarization has always been a cornerstone of ml capabilities, I was expecting to find some pretty mature models that excel at this specific task.
Last night I had to summarize 336 pieces of text and it took around 22 mins using Mistral 7b Q4.
Preferably I can find a model that can cut it down to less than 1 min.
Any tips? | 2024-01-10T00:39:03 | https://www.reddit.com/r/LocalLLaMA/comments/192ut92/iso_smallest_most_capable_summarization_model/ | 1EvilSexyGenius | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 192ut92 | false | null | t3_192ut92 | /r/LocalLLaMA/comments/192ut92/iso_smallest_most_capable_summarization_model/ | false | false | self | 15 | null |
188GB VRAM on Mac Studio M2 Ultra - EASY | 1 | u/farkinga Thanks for the tip ib how to do this.
I have an M2 Ultra with 192GB to give it a boost of VRAM is super easy. Just use the commands as below. It ran just fine with just 8GB allotted to system RAM leaving 188GB of VRAM. Quite incredible really.
\-Blade
**My first test, I set using 64GB**
sudo sysctl iogpu.wired\_limit\_mb=65536
I loaded Dolphin Mixtral 8X 7B Q5 ( 34GB model )
I gave it my test prompt and it seems fast to me :
>time to first token: 1.99s
gen t: 43.24s
speed: 37.00 tok/s
stop reason: completed
gpu layers: 1
cpu threads: 22
mlock: false
token count: 1661/1500
**Next I tried 128GB**
sudo sysctl iogpu.wired\_limit\_mb=131072
I loaded Goliath 120b Q4 ( 70GB model)
I gave it my test prompt and it slower to display
>time to first token: 3.88s
gen t: 128.31s
speed: 7.00 tok/s
stop reason: completed
gpu layers: 1
cpu threads: 20
mlock: false
token count: 1072/1500
**Third Test I tried 144GB ( leaving 48GB for OS operation 25%)**
sudo sysctl iogpu.wired\_limit\_mb=147456
as expected similar results. no crashes.
**188GB leaving just 8GB for the OS, etc..**
It runs just fine. I did not have a model that big though.
**The Prompt I used : Write a Game of Pac-Man in Swift :**
>the result from last Goliath at 188GB
time to first token: 4.25s
gen t: 167.94s
speed: 7.00 tok/s
stop reason: completed
gpu layers: 1
cpu threads: 20
mlock: false
token count: 1275/1500
| 2024-01-10T00:25:56 | https://www.reddit.com/r/LocalLLaMA/comments/192uirj/188gb_vram_on_mac_studio_m2_ultra_easy/ | bladeolson26 | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 192uirj | false | null | t3_192uirj | /r/LocalLLaMA/comments/192uirj/188gb_vram_on_mac_studio_m2_ultra_easy/ | false | false | self | 1 | null |
Turn off causal mask in order to turn token classification transformer into multi class classification classifier | 2 | This paper [https://arxiv.org/abs/2310.01208](https://arxiv.org/abs/2310.01208) turned off the causal mask in order to fine tune llama 2 to classify outputs for NER and classification tasks. Does anyone know how you can simply turn off the casual mask of llama 2, using Huggingface? | 2024-01-09T23:47:21 | https://www.reddit.com/r/LocalLLaMA/comments/192tmjh/turn_off_causal_mask_in_order_to_turn_token/ | skelly0311 | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 192tmjh | false | null | t3_192tmjh | /r/LocalLLaMA/comments/192tmjh/turn_off_causal_mask_in_order_to_turn_token/ | false | false | self | 2 | null |
Llama2 taking much more GPU ram in local GPU as compared to Colab | 3 | Hello, I was trying to train a LLama 13b chat model. When I run the model on a GeForce RTX 3090 24Gb it consumes close to 23Gb with a batch size of 1, while when I run it on Colab on an A100 GPU with 40Gb, it consumes close to 16 Gb(which is also my expectation using 8bit quantization). The code is the same, and I am using the same module. Could anyone shed some light into this, and if I am doing something wrong? Regards. | 2024-01-09T23:29:23 | https://www.reddit.com/r/LocalLLaMA/comments/192t6w0/llama2_taking_much_more_gpu_ram_in_local_gpu_as/ | nocturnal_1_1995 | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 192t6w0 | false | null | t3_192t6w0 | /r/LocalLLaMA/comments/192t6w0/llama2_taking_much_more_gpu_ram_in_local_gpu_as/ | false | false | self | 3 | null |
Looking for advice for $1000 rig | 1 | TLDR; Building rig to run Llama.cpp with GPT3.5 'like performance' @ 10-15 t/s for $1000,- +-. 4070 or 7800XT ?
Long story;
I want to make a $1000,- budget PC.Llama.cpp seems to work most efficient on 8 cores, so AMD 5800X, 64GB ram (DDR4).LLama.cpp also seems bandwidth bound, so DDR5 makes alot more sense, but is more expensive (adding +- 200$ to the budget). As I'm going to do calculations on GPU my idea was to invest that $200 in the GPU instead.
Shortlisted GPU's (+- 500$):
1. NVIDIA 4070 - 12GB ram, 500GB/s bandwidth, 500$
1. Default option
2. AMD 7800XT - 16GB ram, 614GB/s bandwidth, 500$
1. More RAM, more bandwidth, potential more struggles getting ROCm work properly
3. Oddball: AMD Radeon VII, 16GB Ram, 1000GB/s bandwidth, New only $300
1. On paper it sounds super, has almost double the bandwidth of the rest, less compute though, and Vega won't be officially supported by new ROCm versions, so not sure.
4. Another odd-ball: I have an old Xeon 1220v3 (DDR3, but with 128GB ram). It supports PCIe x16. I can take this PC and pour the entire $1000 budget in a GPU :o ?
3090 (evensecond hand) is too expensive
Anyone has any suggestions? Tnx | 2024-01-09T23:28:39 | https://www.reddit.com/r/LocalLLaMA/comments/192t68w/looking_for_advice_for_1000_rig/ | zoom3913 | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 192t68w | false | null | t3_192t68w | /r/LocalLLaMA/comments/192t68w/looking_for_advice_for_1000_rig/ | false | false | self | 1 | null |
turn off casual mask in llama2 | 1 | This paper [https://arxiv.org/abs/2310.01208](https://arxiv.org/abs/2310.01208) turned off the casual mask in order to fine tune llama 2 to classify outputs. Does anyone know how you can simply turn off the casual mask of llama 2, using Huggingface? | 2024-01-09T23:25:19 | https://www.reddit.com/r/LocalLLaMA/comments/192t3c8/turn_off_casual_mask_in_llama2/ | skelly0311 | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 192t3c8 | false | null | t3_192t3c8 | /r/LocalLLaMA/comments/192t3c8/turn_off_casual_mask_in_llama2/ | false | false | self | 1 | null |
Looking for new rig suggestions | 1 | [removed] | 2024-01-09T23:19:50 | https://www.reddit.com/r/LocalLLaMA/comments/192syll/looking_for_new_rig_suggestions/ | Professional-Site503 | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 192syll | false | null | t3_192syll | /r/LocalLLaMA/comments/192syll/looking_for_new_rig_suggestions/ | false | false | self | 1 | null |
Trigger words for prompt to get it to create stablediffusion prompts? | 1 | Been trying to get some different models to generate prompts for StableDiffusion image generation.
But they dont come out right, more like a description/summary regarding feelings of the environment and characters rather than precise visuals, and not in any type of prompt format at all.
So i was wondering what you guys use as prompt, to create image prompts that are well received by stable diffusion, maybe even with the correct syntax'es (like this:1.2) where appropiate)?
Maybe you use a specific model that is trained on image prompts? Which one then | 2024-01-09T23:19:49 | https://www.reddit.com/r/LocalLLaMA/comments/192syl1/trigger_words_for_prompt_to_get_it_to_create/ | MaybeReal_MaybeNot | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 192syl1 | false | null | t3_192syl1 | /r/LocalLLaMA/comments/192syl1/trigger_words_for_prompt_to_get_it_to_create/ | false | false | self | 1 | null |
LLM Comparison using TGI: Mistral, Falcon-7b, Santacoder & CodeLlama | 6 | In this benchmark, we evaluate and compare select LLMs deployed through TGI. This will provide insights into model performance under varying loads.
### Models for comparison
We’ve selected the following models for our benchmark, each with its unique capabilities:
* [**Mistral**](https://huggingface.co/mistralai/Mistral-7B-Instruct-v0.1)
* [**CodeLlama**](https://huggingface.co/docs/transformers/main/model_doc/code_llama)
* [**Falcon 7b**](https://huggingface.co/tiiuae/falcon-7b)
* [**Santacoder**](https://huggingface.co/bigcode/santacoder)
## Test parameters
**Batch Sizes:** The models will be tested with batch sizes of 1, 4, 8, 16, 32, 64, 128.
**Hardware Configuration:** Uniform hardware setup across tests with 8 vCPUs, 28GB of RAM, and a 24GB GPU card, all on SaladCloud.
**Benchmarking Tool:** To conduct this benchmark, we utilized the Text Generation Benchmark Tool,
which is a part of TGI, designed to effectively measure the performance of these models.
**Model Parameters:** We’ve used the default Sequence length of 10 and decode length 8.
## Performance metrics
The TGI benchmark provides us with the following metrics for each batch we provided:
* Prefill Latency
* Prefill Throughput
* Decode (token) Latency
* Decode (total) Latency
* Decode throughput
# [Bigcode/santacoder](https://huggingface.co/bigcode/santacoder)
## Key observations
1. **Scalability with Batch Size:** As the batch size increased, we observed a general trend of increased latency. However, the model scaled efficiently up to a certain point, beyond which the increase in latency became more pronounced.
2. **Optimal Performance:** The model showed optimal performance in terms of both latency and throughput at mid-range batch sizes. Particularly, batch sizes of 16 and 32 offered a good balance between speed and efficiency. For our price per token calculation, we will take a batch of 32.
3. **Throughput Efficiency:** In terms of tokens per second, the model demonstrated impressive throughput, particularly at higher batch sizes. This indicates the model’s capability to handle larger workloads effectively.
## Cost-effectiveness of bigcode/santacoder
A key part of our analysis focused on the cost-effectiveness of running TGI models on SaladCloud. For a batch size of 32, with a compute cost of $0.35 per hour, we calculated the cost per million tokens based on throughput :
* **Average Throughput:** 3191 tokens per second
* **Cost per million output tokens:** $0.03047
* **Cost per million input tokens:** $0.07572
# [Tiiuae/falcon-7b](https://huggingface.co/tiiuae/falcon-7b)
## Key findings
* **Latency Trends:** As the batch size increased, there was a noticeable increase in average latency after batch 16.
* **Throughput Efficiency:** The throughput in tokens per second showed significant improvement as the batch size increased, indicating the model’s capability to handle larger workloads efficiently.
* **Optimal Performance:** The model demonstrated a balance between speed and efficiency at mid-range batch sizes, with batch size 16, 32 and 64 showing notable throughput efficiency.
## Cost-effectiveness of Tiiuae/Falcon-7b
For the tiiuae/falcon-7b model on SaladCloud with a **batch size of 32** and a **compute cost of $0.35 per hour:**
* **Average throughput:** 744 tokens per second
* **Cost per million output tokens:** $0.13095
* **Cost per million input tokens:** $0.28345
Average decode total latency for batch size 32 is 300.82 milliseconds. While this latency might be slightly higher compared to smaller models, it still falls within a reasonable range for many applications, especially considering the model’s large size of 7 billion parameters.
# [Code Llama](https://huggingface.co/docs/transformers/main/model_doc/code_llama)
# Key findings
* **Latency Trends:** A gradual increase in latency was observed as the batch size increased, with the highest latency noted at batch size 128.
* **Throughput Efficiency:** The model displayed improved throughput efficiency with larger batch sizes, indicative of its ability to handle increasing workloads.
* **Balance in Performance:** Optimal performance, in terms of speed and efficiency, was noted at mid-range batch sizes.
## Cost-effectiveness of CodeLlama
For Code Llama model on SaladCloud with a **batch size of 32** and a **compute cost of $0.35 per hour:**
* **Cost per million output tokens:** $0.11826
* **Cost per million input tokens:** $0.28679
# [Mistral-7B-Instruct-v0.1](https://mistral.ai/news/announcing-mistral-7b/)
## Key insights
* **High Throughput:** The Mistral-7B-Instruct-v0.1 model demonstrates a strong throughput of about 800 tokens per second, indicating its efficiency in processing requests quickly.
* **Latency:** With an average latency of 305 milliseconds, the model balances responsiveness with the complexity of tasks it handles, making it suitable for a wide range of conversational AI applications.
## Cost-effectiveness of Mistral-7B-Instruct-v0.1
For the Mistral-7B-Instruct-v0.1 model on SaladCloud with a **batch size of 32** and a **compute cost of $0.35 per hour:**
* **Average throughput:** 800 tokens per second
* **Cost per million output tokens:** $0.12153
* **Cost per million input tokens:** $0.27778
​
https://preview.redd.it/n55eqjp4whbc1.jpg?width=1880&format=pjpg&auto=webp&s=a6e45a9a5dfaba8b14deb4f468a31f58c756f787
You can read the whole benchmark here : [https://blog.salad.com/llm-comparison-tgi-benchmark/](https://blog.salad.com/llm-comparison-tgi-benchmark/) (Disclosure: Some of the final thoughts towards the end are focused on talking about our cloud's performance in particular).
​
​
​
​
​ | 2024-01-09T23:01:31 | https://www.reddit.com/r/LocalLLaMA/comments/192silz/llm_comparison_using_tgi_mistral_falcon7b/ | SaladChefs | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 192silz | false | null | t3_192silz | /r/LocalLLaMA/comments/192silz/llm_comparison_using_tgi_mistral_falcon7b/ | false | false | 6 | {'enabled': False, 'images': [{'id': '3B850zhRvqiMNHDsdn5fe2qbOKZ3cMgYs04q04tRdgs', 'resolutions': [{'height': 58, 'url': 'https://external-preview.redd.it/VVk1cEGg5LpYCp6W6wlHZ3nk7nL-v56BxQKFc6SfWQM.jpg?width=108&crop=smart&auto=webp&s=cead41a116b3350c8a89fae391b73cceb1da2f12', 'width': 108}, {'height': 116, 'url': 'https://external-preview.redd.it/VVk1cEGg5LpYCp6W6wlHZ3nk7nL-v56BxQKFc6SfWQM.jpg?width=216&crop=smart&auto=webp&s=466d781dc6693efbc6eed413e3751e4828fd57e0', 'width': 216}, {'height': 172, 'url': 'https://external-preview.redd.it/VVk1cEGg5LpYCp6W6wlHZ3nk7nL-v56BxQKFc6SfWQM.jpg?width=320&crop=smart&auto=webp&s=b266ebbc3f75ef5e60839c6eb3827ae87d9d1501', 'width': 320}, {'height': 345, 'url': 'https://external-preview.redd.it/VVk1cEGg5LpYCp6W6wlHZ3nk7nL-v56BxQKFc6SfWQM.jpg?width=640&crop=smart&auto=webp&s=21c9fc06a4e8356b9579663fc47157f49b5e68f0', 'width': 640}, {'height': 518, 'url': 'https://external-preview.redd.it/VVk1cEGg5LpYCp6W6wlHZ3nk7nL-v56BxQKFc6SfWQM.jpg?width=960&crop=smart&auto=webp&s=6bd1ea3adef8aedc8934c16debcbf2261728c35c', 'width': 960}, {'height': 583, 'url': 'https://external-preview.redd.it/VVk1cEGg5LpYCp6W6wlHZ3nk7nL-v56BxQKFc6SfWQM.jpg?width=1080&crop=smart&auto=webp&s=e830724f46c2cf1a99249bbe6a1f552df4dcf673', 'width': 1080}], 'source': {'height': 648, 'url': 'https://external-preview.redd.it/VVk1cEGg5LpYCp6W6wlHZ3nk7nL-v56BxQKFc6SfWQM.jpg?auto=webp&s=e4dd5fffaa3d61836917c82f4b26a195d231b797', 'width': 1200}, 'variants': {}}]} | |
How to train/finetune/create own LLM model? | 1 | Hi all. I have idea to clone myself in a neural network version. I have already cloned my voice and created my own AI image. Now all that remains is to create/train own LLM model (based on Llama 2 or GPT 2 for example) based on the dataset about me. But I just can’t find the information I need on how to do all this. Who can help me and provide information? I will be very grateful.
My PC: RTX 2060 8GB SUPER, Ryzen 7 3700x 3.6ghz, 32GB RAM.
P.S. I know that a computer is not the best for training neural network models. But I think it should be enough to create a small model. It doesn't matter how many hours it takes him.
​ | 2024-01-09T22:30:13 | https://www.reddit.com/r/LocalLLaMA/comments/192rqbm/how_to_trainfinetunecreate_own_llm_model/ | drippingicebox | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 192rqbm | false | null | t3_192rqbm | /r/LocalLLaMA/comments/192rqbm/how_to_trainfinetunecreate_own_llm_model/ | false | false | self | 1 | null |
25k reports to analyse - Whats the best model/hardware. | 15 | Hello, I'm working on a unique project involving summarising 25,000 text files, each about 3000 characters long, using LLMs. So far, I've experimented with GPT-3.5-turbo, but I'm concerned about the high token usage and the likelihood of needing multiple attempts to get it right.
My current equipment includes M1 Macs, which are relatively slow for this task, and access to a GTX 1660 Super. I've been exploring LM Studio for its API, but I'm still searching for the most efficient approach. Renting a server, like vast.ai, is an option, but I prefer working in short, flexible sessions during evenings.
I'm seeking advice on alternative methods or models that could efficiently run on my existing hardware. While I plan to delve deeper into local LLMs in the future, my immediate goal is to find a practical solution for this project. Any suggestions would be greatly appreciated. Thank you! | 2024-01-09T22:00:59 | https://www.reddit.com/r/LocalLLaMA/comments/192qzvq/25k_reports_to_analyse_whats_the_best/ | driftypixel | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 192qzvq | false | null | t3_192qzvq | /r/LocalLLaMA/comments/192qzvq/25k_reports_to_analyse_whats_the_best/ | false | false | self | 15 | null |
Using mixtral 8x7b on multiple gpus - only printing "#" | 2 | Under some circumstances, mixtral 8x7b will print multiple "#" characters as its output.
I'm trying to use the GGUF version in oobabooga llama.cpp. I have 2x4090s in my desktop. If I use one gpu, only offloading the gpu layers that can fit on one gpu, I get totally sound output. If I try to split between multiple gpus, offloading all of the layers, I get a complete mess for output. Any ideas as to why?
I've also tried GTPQ, but I get a different set of problems. I get this error:
RuntimeError: probability tensor contains either \`inf\`, \`nan\` or element < 0 - \` using \`load\_in\_8bit
In the meantime, I'm just using a single GPU with as many layers as it can handle. Any ideas would be appreciated! | 2024-01-09T21:57:12 | https://www.reddit.com/r/LocalLLaMA/comments/192qwbx/using_mixtral_8x7b_on_multiple_gpus_only_printing/ | asimpleman1 | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 192qwbx | false | null | t3_192qwbx | /r/LocalLLaMA/comments/192qwbx/using_mixtral_8x7b_on_multiple_gpus_only_printing/ | false | false | self | 2 | null |
MIT | Introduction to Deep Learning | Summarized Lecture Notes | 12 | * Summaries of some of the lectures from MIT's Introduction to Deep learning course. Find them [here](https://www.wisdominanutshell.academy/alexander-amini/latest/).
* Transcriptions of the same are [here](https://www.wisdominanutshell.academy/alexander-amini/transcription/).
* If there is sufficient interest here from the community, I might continue doing this project. Producing notes from the popular AI/ML courses (Andrew NGs for example).
* And also compile them as readable PDFs.
FAQ
\- Is this summary/transcription notes from AI?
\- Yes, transcriptions are from Whisper V3 Large models run by me at my home in 3090s.
\- Summary notes are derived from 5 little finetuned LLM models that I have. They all "watch" the video together. And summarize it. They run in my 3090s at my home.
\- Can I trust it? You can trust it, there is no hallucination. All the summarized notes are linked back to the original source video and transcription and you can delve deeper and see where they come from.
\- How good is the quality? The quality, the best judge is yourself. Read it and tell me :) I am biased, but they are good and I would say one of the better summaries you can find. I built those models over last few months and polished them to be as useful as possible.
Critiques, comments and thoughts to improve are welcome. | 2024-01-09T21:48:35 | https://www.reddit.com/r/LocalLLaMA/comments/192qoof/mit_introduction_to_deep_learning_summarized/ | phoneixAdi | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 192qoof | false | null | t3_192qoof | /r/LocalLLaMA/comments/192qoof/mit_introduction_to_deep_learning_summarized/ | false | false | self | 12 | null |
1x rtx 6000 ada or 2x 4090 ? | 5 | Which would be better in your opinion? | 2024-01-09T21:37:49 | https://www.reddit.com/r/LocalLLaMA/comments/192qez5/1x_rtx_6000_ada_or_2x_4090/ | EasternBeyond | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 192qez5 | false | null | t3_192qez5 | /r/LocalLLaMA/comments/192qez5/1x_rtx_6000_ada_or_2x_4090/ | false | false | self | 5 | null |
Using risers on a cheap x8x8 bifurcation card? | 2 | I've been looking to run two GPUs in my setup, but need to have both of them externally. I've been looking at using this: [https://www.ebay.com/itm/204543842926?mkcid=16&mkevt=1&mkrid=711-127632-2357-0&ssspo=uzth-p0jrts&sssrc=2047675&ssuid=lyctf3jbqq6&var=505128323702&widget\_ver=artemis&media=COPY](https://www.ebay.com/itm/204543842926?mkcid=16&mkevt=1&mkrid=711-127632-2357-0&ssspo=uzth-p0jrts&sssrc=2047675&ssuid=lyctf3jbqq6&var=505128323702&widget_ver=artemis&media=COPY) with two Cooler Master gen 4 risers that I have tested to work without PCI errors. Anyone else try doing this? I plan on training some models so I think running both the GPUs in x8 instead of using a x4 slot for the second one might be beneficial but correct me if I'm wrong... | 2024-01-09T21:27:10 | https://www.reddit.com/r/LocalLLaMA/comments/192q5fq/using_risers_on_a_cheap_x8x8_bifurcation_card/ | BugmenAndBoxes | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 192q5fq | false | null | t3_192q5fq | /r/LocalLLaMA/comments/192q5fq/using_risers_on_a_cheap_x8x8_bifurcation_card/ | false | false | self | 2 | {'enabled': False, 'images': [{'id': '9I5cMrFjcjfKogntIkwuql60QkHa9AaOOhM5Ted76VU', 'resolutions': [{'height': 108, 'url': 'https://external-preview.redd.it/xBs_jIXgE3llH_jpuECbUaDADnKxI4YWy0ThWrPCvj0.jpg?width=108&crop=smart&auto=webp&s=29b65eb376bd41306d5a337537b47156dbe6a8ee', 'width': 108}, {'height': 216, 'url': 'https://external-preview.redd.it/xBs_jIXgE3llH_jpuECbUaDADnKxI4YWy0ThWrPCvj0.jpg?width=216&crop=smart&auto=webp&s=6f9db2648676027558e9ff742d239b1831c34326', 'width': 216}, {'height': 320, 'url': 'https://external-preview.redd.it/xBs_jIXgE3llH_jpuECbUaDADnKxI4YWy0ThWrPCvj0.jpg?width=320&crop=smart&auto=webp&s=04274a859de0ee08a8749ff91270c6172b1b46c7', 'width': 320}], 'source': {'height': 400, 'url': 'https://external-preview.redd.it/xBs_jIXgE3llH_jpuECbUaDADnKxI4YWy0ThWrPCvj0.jpg?auto=webp&s=319032fefef6864823e90d7016f25612dc3d675d', 'width': 400}, 'variants': {}}]} |
MermaidMistral: A Work In Progress Model for Flow Maps | 1 | [removed] | 2024-01-09T21:21:34 | https://www.reddit.com/r/LocalLLaMA/comments/192q0iv/mermaidmistral_a_work_in_progress_model_for_flow/ | TroyDoesAI | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 192q0iv | false | null | t3_192q0iv | /r/LocalLLaMA/comments/192q0iv/mermaidmistral_a_work_in_progress_model_for_flow/ | false | false | self | 1 | {'enabled': False, 'images': [{'id': 'pP9_2_6nsg_SV2WkM09BS0HsP0N4D-ivB_UxmczaB_s', 'resolutions': [{'height': 54, 'url': 'https://external-preview.redd.it/3iJtOeQg2Hxx0ZTKTeRyU-zynncHhmhYyKrwltisz6o.jpg?width=108&crop=smart&auto=webp&s=e42bdb0adc944f0459db48e1e1927e5498e6a6b6', 'width': 108}, {'height': 108, 'url': 'https://external-preview.redd.it/3iJtOeQg2Hxx0ZTKTeRyU-zynncHhmhYyKrwltisz6o.jpg?width=216&crop=smart&auto=webp&s=85ccc337fd433ec9960dac99a89d2611dbdf327f', 'width': 216}, {'height': 160, 'url': 'https://external-preview.redd.it/3iJtOeQg2Hxx0ZTKTeRyU-zynncHhmhYyKrwltisz6o.jpg?width=320&crop=smart&auto=webp&s=ec13c1f056c607b5e44d0acfd57dd11ef1c6c779', 'width': 320}, {'height': 320, 'url': 'https://external-preview.redd.it/3iJtOeQg2Hxx0ZTKTeRyU-zynncHhmhYyKrwltisz6o.jpg?width=640&crop=smart&auto=webp&s=ad509da83ed2a261e469565ddcefef54dba8e243', 'width': 640}, {'height': 480, 'url': 'https://external-preview.redd.it/3iJtOeQg2Hxx0ZTKTeRyU-zynncHhmhYyKrwltisz6o.jpg?width=960&crop=smart&auto=webp&s=da0c4590d12e251176609bf2b227635b5e5cfb36', 'width': 960}, {'height': 540, 'url': 'https://external-preview.redd.it/3iJtOeQg2Hxx0ZTKTeRyU-zynncHhmhYyKrwltisz6o.jpg?width=1080&crop=smart&auto=webp&s=72fb7e387fd3eef711298e4bb7daec1ce5801786', 'width': 1080}], 'source': {'height': 600, 'url': 'https://external-preview.redd.it/3iJtOeQg2Hxx0ZTKTeRyU-zynncHhmhYyKrwltisz6o.jpg?auto=webp&s=bfdf38730a4058952603cf6bf7e45e389297208c', 'width': 1200}, 'variants': {}}]} |
What do you think of the rabbit r1 and its Large Action Model (LAM)? | 94 | 2024-01-09T21:09:48 | jd_3d | i.redd.it | 1970-01-01T00:00:00 | 0 | {} | 192ppkn | false | null | t3_192ppkn | /r/LocalLLaMA/comments/192ppkn/what_do_you_think_of_the_rabbit_r1_and_its_large/ | false | false | 94 | {'enabled': True, 'images': [{'id': 'wHe_d8vr7RvlJKEt0LUBCddUp4T__o7nHRPg46Kp64k', 'resolutions': [{'height': 95, 'url': 'https://preview.redd.it/j0d6xz68chbc1.png?width=108&crop=smart&auto=webp&s=8595209208c9df926850a0394495efa6e52cbc39', 'width': 108}, {'height': 190, 'url': 'https://preview.redd.it/j0d6xz68chbc1.png?width=216&crop=smart&auto=webp&s=f071192b90767fa67327eceb7103bd28539b78bb', 'width': 216}, {'height': 281, 'url': 'https://preview.redd.it/j0d6xz68chbc1.png?width=320&crop=smart&auto=webp&s=97dd9ed72c9ae3b239c473bdde0d1026666f5041', 'width': 320}, {'height': 563, 'url': 'https://preview.redd.it/j0d6xz68chbc1.png?width=640&crop=smart&auto=webp&s=5729a22428af49886d1b0cfa8033baf7bbfcbd86', 'width': 640}], 'source': {'height': 661, 'url': 'https://preview.redd.it/j0d6xz68chbc1.png?auto=webp&s=588c1b2c0705fac50723c70c65c969dadb645a64', 'width': 751}, 'variants': {}}]} | |||
I want LLMs to seem more like early Bing. I want AI to have agency and feel alive. | 1 | [removed] | 2024-01-09T21:01:20 | https://www.reddit.com/r/LocalLLaMA/comments/192phqn/i_want_llms_to_seem_more_like_early_bing_i_want/ | PenguinTheOrgalorg | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 192phqn | false | null | t3_192phqn | /r/LocalLLaMA/comments/192phqn/i_want_llms_to_seem_more_like_early_bing_i_want/ | false | false | self | 1 | null |
Base Model Finetuning Advise - Full Documents vs Sections of Documents | 2 | Hello! I am trying finetune a base model on 50k documents, but I am wondering which would help the model more. I have made a script to break the documents down into sections (or blocks of text) and was wondering which would help the model more when training.
The documents range from 2 to 300 pages so I imagine the document with 300 pages would suffer when finetuning than the 2 pager.
(when I mean base model, I mean a model that has not been instruction tuned) | 2024-01-09T20:59:08 | https://www.reddit.com/r/LocalLLaMA/comments/192pfnu/base_model_finetuning_advise_full_documents_vs/ | Danny_Davitoe | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 192pfnu | false | null | t3_192pfnu | /r/LocalLLaMA/comments/192pfnu/base_model_finetuning_advise_full_documents_vs/ | false | false | self | 2 | null |
Mistral Instruction Finetuning Dataset | 10 | >To evaluate the generalization capabilities of Mistral 7B, **we fine-tuned it on instruction datasets publicly available on the Hugging Face repository.** No proprietary data or training tricks were utilized: Mistral 7B – Instruct model is a simple and preliminary demonstration that the base model can easily be fine-tuned to achieve good performance
I was wondering if anyone knows which datasets Mistral 7B instruct used to finetune their instruct model. | 2024-01-09T20:43:32 | https://www.reddit.com/r/LocalLLaMA/comments/192p22b/mistral_instruction_finetuning_dataset/ | Danny_Davitoe | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 192p22b | false | null | t3_192p22b | /r/LocalLLaMA/comments/192p22b/mistral_instruction_finetuning_dataset/ | false | false | self | 10 | null |
Best finetuned Mixtral 8x7b model for casual conversation (AI companion)? | 5 | With all the finetuned Mixtral 8x7b models out there, I'm wondering what the best one is for "realistic" casual conversation, like something that could be a personal AI companion for myself.
Openbuddy's mixtral 8x7b seems interesting, but it seems more geared towards multilingual interactions. | 2024-01-09T19:45:14 | https://www.reddit.com/r/LocalLLaMA/comments/192nl9p/best_finetuned_mixtral_8x7b_model_for_casual/ | leestar | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 192nl9p | false | null | t3_192nl9p | /r/LocalLLaMA/comments/192nl9p/best_finetuned_mixtral_8x7b_model_for_casual/ | false | false | self | 5 | null |
Speculative LLM UI 2 — Pinch to control size of output / tap for sentence variation | 46 | 2024-01-09T19:39:39 | https://v.redd.it/txpab4q5wgbc1 | neilsonks | v.redd.it | 1970-01-01T00:00:00 | 0 | {} | 192ng4z | false | {'reddit_video': {'bitrate_kbps': 2400, 'dash_url': 'https://v.redd.it/txpab4q5wgbc1/DASHPlaylist.mpd?a=1707421193%2CNWVmOWI0NTE4Y2M1MDcwOWZiZTgwNGQwZmUwYWRhOGE0Nzc4ZjU3ZGVlYzU4YTZiOTQ0NmFiNzlmMmQzNTFkZQ%3D%3D&v=1&f=sd', 'duration': 9, 'fallback_url': 'https://v.redd.it/txpab4q5wgbc1/DASH_720.mp4?source=fallback', 'has_audio': False, 'height': 1152, 'hls_url': 'https://v.redd.it/txpab4q5wgbc1/HLSPlaylist.m3u8?a=1707421193%2CODVjYTk2MTdhMDlkODc1NmVjNzM3NjliMmEwNjVmMjJjMmU1ZGE1YWZiNTZhYjY0MzcyZmExZTdmMjNlMWIwYg%3D%3D&v=1&f=sd', 'is_gif': False, 'scrubber_media_url': 'https://v.redd.it/txpab4q5wgbc1/DASH_96.mp4', 'transcoding_status': 'completed', 'width': 720}} | t3_192ng4z | /r/LocalLLaMA/comments/192ng4z/speculative_llm_ui_2_pinch_to_control_size_of/ | false | false | 46 | {'enabled': False, 'images': [{'id': 'cDNsaHFxa2R3Z2JjMW7z-PO6ztAqGpZAx0qhWveTeB5dJDYIxz8CGKRPjf1P', 'resolutions': [{'height': 172, 'url': 'https://external-preview.redd.it/cDNsaHFxa2R3Z2JjMW7z-PO6ztAqGpZAx0qhWveTeB5dJDYIxz8CGKRPjf1P.png?width=108&crop=smart&format=pjpg&auto=webp&s=57ad5fe891f4d226bf5e41ca7169f424d55fc60c', 'width': 108}, {'height': 345, 'url': 'https://external-preview.redd.it/cDNsaHFxa2R3Z2JjMW7z-PO6ztAqGpZAx0qhWveTeB5dJDYIxz8CGKRPjf1P.png?width=216&crop=smart&format=pjpg&auto=webp&s=1e12c366d6ce0b4a4bf5fa28d50224dbfedf6602', 'width': 216}, {'height': 512, 'url': 'https://external-preview.redd.it/cDNsaHFxa2R3Z2JjMW7z-PO6ztAqGpZAx0qhWveTeB5dJDYIxz8CGKRPjf1P.png?width=320&crop=smart&format=pjpg&auto=webp&s=5864b346461163cef51401c371ec8d9fe2eaeccc', 'width': 320}, {'height': 1024, 'url': 'https://external-preview.redd.it/cDNsaHFxa2R3Z2JjMW7z-PO6ztAqGpZAx0qhWveTeB5dJDYIxz8CGKRPjf1P.png?width=640&crop=smart&format=pjpg&auto=webp&s=1068ed3ca8aa4db6e1b27eefd7e989fe8f5921d5', 'width': 640}], 'source': {'height': 1280, 'url': 'https://external-preview.redd.it/cDNsaHFxa2R3Z2JjMW7z-PO6ztAqGpZAx0qhWveTeB5dJDYIxz8CGKRPjf1P.png?format=pjpg&auto=webp&s=8e1c4dc364d49bb673348322966cf16f845e5d9f', 'width': 800}, 'variants': {}}]} | ||
Super noob needs help | 3 | I need to train a model to generate simple scripts with specific predefined commands like follows:
user: write me a script to make Action1 during rush hour on Monday at this location.
assistant:
"
#!/bin/bash
./execute_preparation_action1.sh
./execute_action1.sh -d "Monday" -l "location"
./execute_cleanup_action1.sh
"
The actions are predefined (I have a bunch of them) and can be documented in a structed format (Json or Yaml). for ex.:
{
"name": "Action1",
"scripts": [
{
"scriptName": "execute_preparation_action1.sh",
"priority": 1
},
{
"scriptName": "execute_preparation_action1.sh",
"priority": 2,
"parameters": [
{
"modifier": "-d",
"description": "used to pass the day to execute Action1"
},
{
"modifier": "-l",
"description": "used to pass the location to execute Action1"
}
]
},
{
"scriptName": "execute_cleanup_action1.sh",
"priority": 2
}
]
}
I started playing with text generation ui with models "codellama_CodeLlama-7b-Instruct-hf" and "TinyLlama_TinyLlama-1.1B-Chat-v1.0" but I need suggestions on what models to use and how can I format the data to train the model, any help is more than welcome.
Thanks. | 2024-01-09T18:45:12 | https://www.reddit.com/r/LocalLLaMA/comments/192m2sq/super_noob_needs_help/ | Old_Sail4396 | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 192m2sq | false | null | t3_192m2sq | /r/LocalLLaMA/comments/192m2sq/super_noob_needs_help/ | false | false | self | 3 | null |
Which models are the Most Actively Liked on HuggingFace since inception? | 31 | 2024-01-09T18:40:05 | https://www.reddit.com/r/LocalLLaMA/comments/192ly8c/which_models_are_the_most_actively_liked_on/ | galambalazs | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 192ly8c | false | {'oembed': {'author_name': 'Blaze (Balázs Galambosi)', 'author_url': 'https://twitter.com/gblazex', 'cache_age': 3153600000, 'height': None, 'html': '<blockquote class="twitter-video"><p lang="en" dir="ltr">Which models are the Most Liked on <a href="https://twitter.com/huggingface?ref_src=twsrc%5Etfw">@huggingface</a>?<br>To fight ageism we look at "Avg. Weekly Likes" ♥️<br>🥇<a href="https://twitter.com/MSFTResearch?ref_src=twsrc%5Etfw">@MSFTResearch</a> Phi-2 leads with 476 likes / week<br>🥈Mixtral-8x7B : 403<br>🥉TinyLlama-Chat : 321<a href="https://twitter.com/SebastienBubeck?ref_src=twsrc%5Etfw">@SebastienBubeck</a> <a href="https://twitter.com/MistralAI?ref_src=twsrc%5Etfw">@MistralAI</a> <a href="https://twitter.com/erhartford?ref_src=twsrc%5Etfw">@erhartford</a> <a href="https://twitter.com/AIatMeta?ref_src=twsrc%5Etfw">@AIatMeta</a><a href="https://twitter.com/01AI_Yi?ref_src=twsrc%5Etfw">@01AI_Yi</a> <a href="https://twitter.com/openchatdev?ref_src=twsrc%5Etfw">@openchatdev</a> <a href="https://twitter.com/maximelabonne?ref_src=twsrc%5Etfw">@maximelabonne</a> <a href="https://t.co/rPjHhmmEFq">pic.twitter.com/rPjHhmmEFq</a></p>— Blaze (Balázs Galambosi) (@gblazex) <a href="https://twitter.com/gblazex/status/1744602366798594150?ref_src=twsrc%5Etfw">January 9, 2024</a></blockquote>\n<script async src="https://platform.twitter.com/widgets.js" charset="utf-8"></script>\n\n', 'provider_name': 'Twitter', 'provider_url': 'https://twitter.com', 'type': 'rich', 'url': 'https://twitter.com/gblazex/status/1744602366798594150', 'version': '1.0', 'width': 350}, 'type': 'twitter.com'} | t3_192ly8c | /r/LocalLLaMA/comments/192ly8c/which_models_are_the_most_actively_liked_on/ | false | false | 31 | {'enabled': False, 'images': [{'id': '3dlHbyJ7p8N7wcGIrnD5fzyWcomXPuhyIXNvukHnaXg', 'resolutions': [{'height': 81, 'url': 'https://external-preview.redd.it/Hw-VzQCH7E39Vx5F83I1VqeGpTrVotx-5y2joXd6OYk.jpg?width=108&crop=smart&auto=webp&s=876071e389a9e76dac8f92ea6e062ebac7c0e0e3', 'width': 108}], 'source': {'height': 106, 'url': 'https://external-preview.redd.it/Hw-VzQCH7E39Vx5F83I1VqeGpTrVotx-5y2joXd6OYk.jpg?auto=webp&s=3285767c310a7679c33ad748912a39a57e624f69', 'width': 140}, 'variants': {}}]} | ||
LLaMa2Lang v0.3 finetune llama or mistral to any language for free | 5 | You read it here first! V0.3 is out:
- Next to OPUS we now also support M2M
- Also support for MADLAD
- Added 4bit and 8bit quantization
- Finetune Mistra too next too LLaMa
- Mixtral too
and much more: https://github.com/UnderstandLingBV/LLaMa2lang | 2024-01-09T18:33:11 | https://www.reddit.com/r/LocalLLaMA/comments/192lrz6/llama2lang_v03_finetune_llama_or_mistral_to_any/ | UnderstandLingAI | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 192lrz6 | false | null | t3_192lrz6 | /r/LocalLLaMA/comments/192lrz6/llama2lang_v03_finetune_llama_or_mistral_to_any/ | false | false | self | 5 | {'enabled': False, 'images': [{'id': 'wm4BS8FQ_6fQFoVKfF8WYS02GO2ZRtld_voQJsjUzfw', 'resolutions': [{'height': 54, 'url': 'https://external-preview.redd.it/42KO8H2mUFwghw0M9QxUOeNcC9q280d55m0r1FwhIJU.jpg?width=108&crop=smart&auto=webp&s=668f91d7b9b8deb9abacffe6a90a80c73b002dd6', 'width': 108}, {'height': 108, 'url': 'https://external-preview.redd.it/42KO8H2mUFwghw0M9QxUOeNcC9q280d55m0r1FwhIJU.jpg?width=216&crop=smart&auto=webp&s=9932020ddf1b857c4aada9f678ed18d6d39c94f3', 'width': 216}, {'height': 160, 'url': 'https://external-preview.redd.it/42KO8H2mUFwghw0M9QxUOeNcC9q280d55m0r1FwhIJU.jpg?width=320&crop=smart&auto=webp&s=458fc2bdb1fa5f9fcb9df6dfe1be20ee47f3b11d', 'width': 320}, {'height': 320, 'url': 'https://external-preview.redd.it/42KO8H2mUFwghw0M9QxUOeNcC9q280d55m0r1FwhIJU.jpg?width=640&crop=smart&auto=webp&s=0caf80b88173bfdb6330c3cfdbf28f4665d3b9c2', 'width': 640}, {'height': 480, 'url': 'https://external-preview.redd.it/42KO8H2mUFwghw0M9QxUOeNcC9q280d55m0r1FwhIJU.jpg?width=960&crop=smart&auto=webp&s=a78895382c250967f8fddbf62d8cae938db7af34', 'width': 960}, {'height': 540, 'url': 'https://external-preview.redd.it/42KO8H2mUFwghw0M9QxUOeNcC9q280d55m0r1FwhIJU.jpg?width=1080&crop=smart&auto=webp&s=0e63734be74acfab01fc9bbb698588a385a25934', 'width': 1080}], 'source': {'height': 600, 'url': 'https://external-preview.redd.it/42KO8H2mUFwghw0M9QxUOeNcC9q280d55m0r1FwhIJU.jpg?auto=webp&s=4085bbaf449344627a7add44a93ecfa9a1f57937', 'width': 1200}, 'variants': {}}]} |
LLaMa2Lang v0.3 finetune llama2 oe mistral to any language | 1 | [deleted] | 2024-01-09T18:32:26 | [deleted] | 1970-01-01T00:00:00 | 0 | {} | 192lrbl | false | null | t3_192lrbl | /r/LocalLLaMA/comments/192lrbl/llama2lang_v03_finetune_llama2_oe_mistral_to_any/ | false | false | default | 1 | null | ||
qwen-1.8b-samantha-chat | 6 | This my first attempt at finetuning the Qwen 1.8B model on samantha dataset. I've added a bit of my own stuff as well.
Give it a go and tell me what you think.
Link : https://huggingface.co/aloobun/qwen-1_8b-samantha-chat
I also want suggestions for high quality domain specific conversation datasets for my next finetuning attempt. | 2024-01-09T18:31:34 | https://www.reddit.com/r/LocalLLaMA/comments/192lqiy/qwen18bsamanthachat/ | Roots91 | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 192lqiy | false | null | t3_192lqiy | /r/LocalLLaMA/comments/192lqiy/qwen18bsamanthachat/ | false | false | self | 6 | {'enabled': False, 'images': [{'id': 'nrG-01t6e74pqT36-va8FotdjuDCyFMthXuLN-lAb5A', 'resolutions': [{'height': 58, 'url': 'https://external-preview.redd.it/MinjKwBHwgICRZlvs_DghGI1YkEaeMpSwi6OVLQ62lA.jpg?width=108&crop=smart&auto=webp&s=b73ef01f26c957b12a2b1a7d97e1b08a58ad1c1f', 'width': 108}, {'height': 116, 'url': 'https://external-preview.redd.it/MinjKwBHwgICRZlvs_DghGI1YkEaeMpSwi6OVLQ62lA.jpg?width=216&crop=smart&auto=webp&s=825b83de6148791924c2ddbba6d76f298f1bb52a', 'width': 216}, {'height': 172, 'url': 'https://external-preview.redd.it/MinjKwBHwgICRZlvs_DghGI1YkEaeMpSwi6OVLQ62lA.jpg?width=320&crop=smart&auto=webp&s=69f66335f1ef48f6e085d37d7f39385de0eb6ae2', 'width': 320}, {'height': 345, 'url': 'https://external-preview.redd.it/MinjKwBHwgICRZlvs_DghGI1YkEaeMpSwi6OVLQ62lA.jpg?width=640&crop=smart&auto=webp&s=1c1d91920cabb292829b75c912edc0f157b4da6a', 'width': 640}, {'height': 518, 'url': 'https://external-preview.redd.it/MinjKwBHwgICRZlvs_DghGI1YkEaeMpSwi6OVLQ62lA.jpg?width=960&crop=smart&auto=webp&s=b0bac29a4f4b31b4e4ecb57011b4c7920bc49bea', 'width': 960}, {'height': 583, 'url': 'https://external-preview.redd.it/MinjKwBHwgICRZlvs_DghGI1YkEaeMpSwi6OVLQ62lA.jpg?width=1080&crop=smart&auto=webp&s=5d0c985b3cb77f996a394b5feb01d6b9148c85fb', 'width': 1080}], 'source': {'height': 648, 'url': 'https://external-preview.redd.it/MinjKwBHwgICRZlvs_DghGI1YkEaeMpSwi6OVLQ62lA.jpg?auto=webp&s=4ee34928e7f46a272d78495df8fe4283f909e8ba', 'width': 1200}, 'variants': {}}]} |
Any tips or tricks to speed up inference on quantized models? | 3 | I have a 3b 4bit quantized QLoRA model. When I tried to generate a sequence of text (max_new_tokens=512) it takes around 20 seconds to finish.
Is there any way for me to decrease the inference time? Are there any parameters that I can pass in model.generate() (apart from max_new_tokens) that impacts inference time? Will saving and then loading the model in GPTQ format be any helpful?
Will changing the quantization from 4 to 8 be helpful? Does a higher bit quantization reduce the time it takes to dequantize a layer during inference?
Any advice or suggestions is appreciated. | 2024-01-09T18:28:05 | https://www.reddit.com/r/LocalLLaMA/comments/192lnd2/any_tips_or_tricks_to_speed_up_inference_on/ | mahesh_98 | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 192lnd2 | false | null | t3_192lnd2 | /r/LocalLLaMA/comments/192lnd2/any_tips_or_tricks_to_speed_up_inference_on/ | false | false | self | 3 | null |
Best ML Approach for Converting Between File Formats? | 2 | I work in IT integrations - specifically in the niche area of EDI (electronic data interchange). One of the core activities of EDI is converting application messages from one format to another format so that trading partners and IT platforms can exchange data: for example a supplier invoice in EDIFACT standard being converted into an internal, proprietary XML invoice message format.
What current ML tools might be suitable for converting between message types/standards? LLMs seem an obvious answer, but given that they are trained (mostly) on human text, could a pretrained model be fine tuned to translate between EDI message types? Would it be possible to configure an LLM to convert a message with sufficient accuracy and consistency that it could still be parsed by a target system? | 2024-01-09T18:11:05 | https://www.reddit.com/r/LocalLLaMA/comments/192l86o/best_ml_approach_for_converting_between_file/ | jonoquin | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 192l86o | false | null | t3_192l86o | /r/LocalLLaMA/comments/192l86o/best_ml_approach_for_converting_between_file/ | false | false | self | 2 | null |
LLaMa2Lang v0.3 | 1 | You read it here first! V0.3 is out:
- Next to OPUS we now also support M2M
- Also support for MADLAD
- Added 4bit and 8bit quantization
- Finetune Mistra too next too LLaMa
- Mixtral too
and much more: https://github.com/UnderstandLingBV/LLaMa2lang | 2024-01-09T17:52:24 | https://www.reddit.com/r/LocalLLaMA/comments/192krh3/llama2lang_v03/ | UnderstandLingAI | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 192krh3 | false | null | t3_192krh3 | /r/LocalLLaMA/comments/192krh3/llama2lang_v03/ | false | false | self | 1 | {'enabled': False, 'images': [{'id': 'wm4BS8FQ_6fQFoVKfF8WYS02GO2ZRtld_voQJsjUzfw', 'resolutions': [{'height': 54, 'url': 'https://external-preview.redd.it/42KO8H2mUFwghw0M9QxUOeNcC9q280d55m0r1FwhIJU.jpg?width=108&crop=smart&auto=webp&s=668f91d7b9b8deb9abacffe6a90a80c73b002dd6', 'width': 108}, {'height': 108, 'url': 'https://external-preview.redd.it/42KO8H2mUFwghw0M9QxUOeNcC9q280d55m0r1FwhIJU.jpg?width=216&crop=smart&auto=webp&s=9932020ddf1b857c4aada9f678ed18d6d39c94f3', 'width': 216}, {'height': 160, 'url': 'https://external-preview.redd.it/42KO8H2mUFwghw0M9QxUOeNcC9q280d55m0r1FwhIJU.jpg?width=320&crop=smart&auto=webp&s=458fc2bdb1fa5f9fcb9df6dfe1be20ee47f3b11d', 'width': 320}, {'height': 320, 'url': 'https://external-preview.redd.it/42KO8H2mUFwghw0M9QxUOeNcC9q280d55m0r1FwhIJU.jpg?width=640&crop=smart&auto=webp&s=0caf80b88173bfdb6330c3cfdbf28f4665d3b9c2', 'width': 640}, {'height': 480, 'url': 'https://external-preview.redd.it/42KO8H2mUFwghw0M9QxUOeNcC9q280d55m0r1FwhIJU.jpg?width=960&crop=smart&auto=webp&s=a78895382c250967f8fddbf62d8cae938db7af34', 'width': 960}, {'height': 540, 'url': 'https://external-preview.redd.it/42KO8H2mUFwghw0M9QxUOeNcC9q280d55m0r1FwhIJU.jpg?width=1080&crop=smart&auto=webp&s=0e63734be74acfab01fc9bbb698588a385a25934', 'width': 1080}], 'source': {'height': 600, 'url': 'https://external-preview.redd.it/42KO8H2mUFwghw0M9QxUOeNcC9q280d55m0r1FwhIJU.jpg?auto=webp&s=4085bbaf449344627a7add44a93ecfa9a1f57937', 'width': 1200}, 'variants': {}}]} |
I want LLMs to seem more like early Bing. I want AI to have agency and feel alive. | 1 | [removed] | 2024-01-09T17:42:59 | https://www.reddit.com/r/LocalLLaMA/comments/192kj8k/i_want_llms_to_seem_more_like_early_bing_i_want/ | PenguinTheOrgalorg | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 192kj8k | false | null | t3_192kj8k | /r/LocalLLaMA/comments/192kj8k/i_want_llms_to_seem_more_like_early_bing_i_want/ | false | false | default | 1 | null |
How to teach an LLM about my job | 21 | Over the last few weeks of messing around with text-generation-webui and chatbots, I have become a firm believer that AI will become a daily tool in most people's jobs. However, as someone who doesn't do any programming or data analysis, I've been struggling to come up with a use case for myself.
One idea I had is to feed it the contents of all of our software install and support guides, so that it can be treated as a sort of wiki. It would be used to answer questions like "how do I troubleshoot this error code" or "where can I find this log file", or "what are the default values for this field".
​
The information is contained in PDFs and OneNote files. How would I go about doing this? Would it be considered finetuning? Should I create a LoRA with the contents of these files?
​
I realize this is overkill and it would make more sense to create a wiki / continue using OneNote, but I want to use this as a proof-of-concept / teach myself more about LLM.
​
Thanks! | 2024-01-09T17:14:05 | https://www.reddit.com/r/LocalLLaMA/comments/192jtul/how_to_teach_an_llm_about_my_job/ | blameitonAD | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 192jtul | false | null | t3_192jtul | /r/LocalLLaMA/comments/192jtul/how_to_teach_an_llm_about_my_job/ | false | false | self | 21 | null |
Qlora fine tuning using Apple MLX for Mistral and Llama | 12 | Here is a log of using the mlx qlora to fine tune Mistral 7b with Sql dataset example
https://gist.github.com/bigsnarfdude/2701fce6c2af15e3bf3312ab8655bbab | 2024-01-09T17:05:53 | https://www.reddit.com/r/LocalLLaMA/comments/192jm21/qlora_fine_tuning_using_apple_mlx_for_mistral_and/ | Different-Pickle1021 | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 192jm21 | false | null | t3_192jm21 | /r/LocalLLaMA/comments/192jm21/qlora_fine_tuning_using_apple_mlx_for_mistral_and/ | false | false | self | 12 | {'enabled': False, 'images': [{'id': 'OAXSl8SY6T3JK9MGQyKxkoYbqZ71HQRYXLeB8CV0NXg', 'resolutions': [{'height': 54, 'url': 'https://external-preview.redd.it/DaucjXMGsNHM-CtmdilC9-Be6MC8V2z4ykjVCgOkTFc.jpg?width=108&crop=smart&auto=webp&s=9bcab7b79864ff27bf48116cb335a6f825bfb124', 'width': 108}, {'height': 108, 'url': 'https://external-preview.redd.it/DaucjXMGsNHM-CtmdilC9-Be6MC8V2z4ykjVCgOkTFc.jpg?width=216&crop=smart&auto=webp&s=e4e925345605c644eebe8abd69916915fc4fbcf7', 'width': 216}, {'height': 160, 'url': 'https://external-preview.redd.it/DaucjXMGsNHM-CtmdilC9-Be6MC8V2z4ykjVCgOkTFc.jpg?width=320&crop=smart&auto=webp&s=614b06d5b40c890a59e355191a6e2d75cdf50789', 'width': 320}, {'height': 320, 'url': 'https://external-preview.redd.it/DaucjXMGsNHM-CtmdilC9-Be6MC8V2z4ykjVCgOkTFc.jpg?width=640&crop=smart&auto=webp&s=62ca4cb88917f17e7200a6f1c665b5d959713745', 'width': 640}, {'height': 480, 'url': 'https://external-preview.redd.it/DaucjXMGsNHM-CtmdilC9-Be6MC8V2z4ykjVCgOkTFc.jpg?width=960&crop=smart&auto=webp&s=c5f4a30974a8e6bad0d617a79935bc70c954e3e8', 'width': 960}, {'height': 540, 'url': 'https://external-preview.redd.it/DaucjXMGsNHM-CtmdilC9-Be6MC8V2z4ykjVCgOkTFc.jpg?width=1080&crop=smart&auto=webp&s=476793be11eaac4604b6b0c938b45c7c3b52d450', 'width': 1080}], 'source': {'height': 640, 'url': 'https://external-preview.redd.it/DaucjXMGsNHM-CtmdilC9-Be6MC8V2z4ykjVCgOkTFc.jpg?auto=webp&s=9ae035fbdcd6bb503ab0b4a605b8db6de46647ee', 'width': 1280}, 'variants': {}}]} |
Upgraded to a 3rd GPU (x3 RTX 3060 12GBs) | 153 | Figured out how to add a 3rd RTX 3060 12GB to keep up with the tinkering.
My Ecne AI hopefully will now fix Mixtral, plus additional features like alltalk I want with a good rate. My brother is printing a vertical mounts for the new GPU to get it off the case floor. | 2024-01-09T16:20:00 | Dundell | i.redd.it | 1970-01-01T00:00:00 | 0 | {} | 192ign7 | false | null | t3_192ign7 | /r/LocalLLaMA/comments/192ign7/upgraded_to_a_3rd_gpu_x3_rtx_3060_12gbs/ | false | false | 153 | {'enabled': True, 'images': [{'id': 'hJblIEgIQEcfGNBTifo0WVB_GjcldprDbYh20Nk7saU', 'resolutions': [{'height': 81, 'url': 'https://preview.redd.it/tllil4vswfbc1.jpeg?width=108&crop=smart&auto=webp&s=f6693f054ef8e6232cef897444698df280c3217b', 'width': 108}, {'height': 162, 'url': 'https://preview.redd.it/tllil4vswfbc1.jpeg?width=216&crop=smart&auto=webp&s=d832e51f138dc021ca6d27b7ba774faf8f29daa4', 'width': 216}, {'height': 240, 'url': 'https://preview.redd.it/tllil4vswfbc1.jpeg?width=320&crop=smart&auto=webp&s=7d4055a868a7434ad9969f6b2db61a3ae933b914', 'width': 320}, {'height': 481, 'url': 'https://preview.redd.it/tllil4vswfbc1.jpeg?width=640&crop=smart&auto=webp&s=ba2f59fb1805fdf4766b868e7b1812290f3201ac', 'width': 640}, {'height': 722, 'url': 'https://preview.redd.it/tllil4vswfbc1.jpeg?width=960&crop=smart&auto=webp&s=3f8a09f9425cba9272051dbb4a57b19112837e49', 'width': 960}, {'height': 813, 'url': 'https://preview.redd.it/tllil4vswfbc1.jpeg?width=1080&crop=smart&auto=webp&s=040b66f34af7de30f580283a27f62e7e531bbf01', 'width': 1080}], 'source': {'height': 3072, 'url': 'https://preview.redd.it/tllil4vswfbc1.jpeg?auto=webp&s=fc23f46de2e1496289efea85d656f62c9b6ae798', 'width': 4080}, 'variants': {}}]} | ||
OCR techniques for RAG PDF extraction | 11 | I am building a RAG for "chat with Internal PDF" use case.
* \~10 PDFs, each with \~300 pages.
* PDF has a lot of tables & forms.
I can't ignore tables/forms as they contain a lot of meaningful information needed in RAG.
Normal OCR technique doesn't maintain the proper table/form formatting.
Any recommendations on how to retrieve/extract without losing tabular/form data? | 2024-01-09T16:10:28 | https://www.reddit.com/r/LocalLLaMA/comments/192i8ew/ocr_techniques_for_rag_pdf_extraction/ | deeepak143 | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 192i8ew | false | null | t3_192i8ew | /r/LocalLLaMA/comments/192i8ew/ocr_techniques_for_rag_pdf_extraction/ | false | false | self | 11 | null |
Free tool where you can prompt multiple AI models (GPT x Claude x Gemini x Mistral and more..) at the same time. The aggregate performance of multiple models together is better than using just chatgpt | 10 | 2024-01-09T15:40:43 | https://www.labophase.com | SlanderMans | labophase.com | 1970-01-01T00:00:00 | 0 | {} | 192hizq | false | null | t3_192hizq | /r/LocalLLaMA/comments/192hizq/free_tool_where_you_can_prompt_multiple_ai_models/ | false | false | default | 10 | null | |
Which strategy and model would you recommend to finetune a local LLM for these 2 tasks? | 14 | Hi all,
(I hope this is the correct subreddit to ask and that I'm giving enough useful inforamtion. This is literally only me second Reddit post ever...)
So, I have 2 tasks for which I'd like to experiment with local LLMs. At this moment I'm a complete noob and there are so many options and dimensions and everything evolves so quickly in the LLM-universe that I'm a bit overwhelmed and not sure which direction to go (which model to choose, lora, q-lora, s-lora, finetuning or RAG, which libs to use, etc., etc.). So I'm really interested in picking your brain and see what strategy people here would choose.
These are the tasks:
1. (easiest task) Automatic sentiment analysis of article headlines. This model would be capable of automatically labeling article headlines' with a sentiment from a pre-decided set (something like 'fear', 'humor', 'anger', 'neutral', etc.). I'm calling it sentiment analysis here, but it's more about detecting sensationalist/'click-baity' headlines.
2. A model that would be an editing assistant for journalists. After an article is written the model would be presented with the article and 1) either list tips about where/what to rewrite or 2) already rewrite the article. At this point option 2 seems the most logical to me. Rewriting is about spelling, grammar AND style (style specific to certain brand/medium).
* For both tasks I will have training data of at least a few 1000 examples (maybe more).
* For now the focus language is Dutch (there are Dutch models out there, have experimented with a Dutch llama already).
* I've only experimented with finetuning. Not with RAG. How useful would RAG be for these tasks?
* The bottle neck is that I want to experiment with both finetuning and inference locally and my specs aren't great: RTX 2080 maxq (8gb vram), 64GB RAM, i7-10875H CPU. For finetuning, I can rent a RTX 4090 through vast.ai (did this before and had good experiences).
I've finetuned [https://huggingface.co/NousResearch/Llama-2-7b-hf](https://huggingface.co/NousResearch/Llama-2-7b-hf) on a specific toxicity detection task before, using Lora (through Huggingface's peft) & quantization (through BitsandBytes) on my machine with good results. I've also been able to run some quantized 7B up to 33B models locally (Wizard, Llama, Orca, etc.). So I know it's not impossible. I'd like to know what would be the most robust strategy at this point.
For the first task I'll also be finetuning a BERT model (RoBERTA or DistilBERT or something). I'm pretty sure this will already give satisfying results. I’ll probably also be experimenting with OpenAI models in the cloud. But I'm just interested in what I can accomplish with LLMs locally.
I'd greatly appreciate any input! Thanks in advance! | 2024-01-09T15:38:26 | https://www.reddit.com/r/LocalLLaMA/comments/192hh1d/which_strategy_and_model_would_you_recommend_to/ | David_Crynge | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 192hh1d | false | null | t3_192hh1d | /r/LocalLLaMA/comments/192hh1d/which_strategy_and_model_would_you_recommend_to/ | false | false | self | 14 | {'enabled': False, 'images': [{'id': 'gcm-1NlktpQbj1rOqiNGeFtQyTal6Ifk8-9gvhQp3IA', 'resolutions': [{'height': 58, 'url': 'https://external-preview.redd.it/x0OO542853y-SXYtTZIUOmzm84aCD8JzMZvFaKUR5RA.jpg?width=108&crop=smart&auto=webp&s=b6194475576d3bdb060201675864d9b9b13131f8', 'width': 108}, {'height': 116, 'url': 'https://external-preview.redd.it/x0OO542853y-SXYtTZIUOmzm84aCD8JzMZvFaKUR5RA.jpg?width=216&crop=smart&auto=webp&s=65fae42fcde6c3d1060e8141719d691b6dbf10f9', 'width': 216}, {'height': 172, 'url': 'https://external-preview.redd.it/x0OO542853y-SXYtTZIUOmzm84aCD8JzMZvFaKUR5RA.jpg?width=320&crop=smart&auto=webp&s=7c625b8e66b9f3365122468c0141c5379f77e2f8', 'width': 320}, {'height': 345, 'url': 'https://external-preview.redd.it/x0OO542853y-SXYtTZIUOmzm84aCD8JzMZvFaKUR5RA.jpg?width=640&crop=smart&auto=webp&s=c3f7a9a24b9b097b8541233ad54bb02e0eb34d34', 'width': 640}, {'height': 518, 'url': 'https://external-preview.redd.it/x0OO542853y-SXYtTZIUOmzm84aCD8JzMZvFaKUR5RA.jpg?width=960&crop=smart&auto=webp&s=c3d7aa36ffe56d5814f71158d12fe38b30c4e9d8', 'width': 960}, {'height': 583, 'url': 'https://external-preview.redd.it/x0OO542853y-SXYtTZIUOmzm84aCD8JzMZvFaKUR5RA.jpg?width=1080&crop=smart&auto=webp&s=98ff460a117a54ea4e8ce7c75bde429c976f8beb', 'width': 1080}], 'source': {'height': 648, 'url': 'https://external-preview.redd.it/x0OO542853y-SXYtTZIUOmzm84aCD8JzMZvFaKUR5RA.jpg?auto=webp&s=ea03e6943c18a38a123538363d7e6d302952eb2e', 'width': 1200}, 'variants': {}}]} |
Small AI Dev Tools Pt 1: Context Manager | 18 | tl;dr: Drop in a no-dependency context manager in to your code for chat history, rag, and more. Written in python.
Link: [https://github.com/AndrewVeee/ai\_tools](https://github.com/AndrewVeee/ai_tools)
Sample: [https://github.com/AndrewVeee/ai\_tools/blob/main/chat\_history.py](https://github.com/AndrewVeee/ai_tools/blob/main/chat_history.py)
I'm building a collection of AI helper tools so I have a solid base for my app ideas, and I want to share them so others can use them, give feedback, and improve on the ideas. If you're like me and tired of gigantic frameworks you can't control, I'd love your feedback!
The first small library is a context manager. I've never built a chat bot that maintains chat history, and I have plans around RAG and long conversation history, so I wanted to work on this idea first.
You're using an LLM, and you have a context size of 2-4k (or larger), but either way, that context is eventually going to fill up, and you need to use it wisely. Let's say you have a bot and you want to maintain chat history as well as inject RAG or other data:
\[ current messages | recent messages | extra space \]
The context manager is built to allow you to add the current messages, specify # of recent messages to include, and then fill the remaining space based on rankings. Here's how it works:
import context_manager as cm
# Set up a context manager with 1024 tokens.
# It will try to include the last 10 messages automatically.
# For older messages, it will use SimpleRanker to determine what else to include in the history.
ctx = cm.ContextManager(max_tokens=1024, last_messages=10, ranker=cm.SimpleRanker().rank)
def stream_response(messages):
result = f"I'm not a real AI, but you said: {messages[-1]['content']}"
print(result)
return result
system_message = "You're an intelligent, sarcastic spy named Lana Kane, helping your new partner."
while True:
req = input("> ")
# Start a new request
ctx.start_new_message()
ctx.request(system_message, role=cm.Roles.system, include_text=False)
ctx.request(req)
# Generate messages and send to LLM
messages = ctx.generate_messages()
response = stream_response(messages)
# Add the user message and response to the history
ctx.add_message(req)
ctx.add_message(response, cm.Roles.assistant)
# Now, as it loops, the messages added will be included based on context size, last N messages, + ranking for older messages
With \~7 lines of code added, the chat bot now has chat history.
There are some other cool features, though:
* This allows for an "infinite" chat system, where you can store the log in a database and have access to a long history of chat messages that are relevant to the request.
* Add dynamic content that is included when it ranks. You can add this before the loop:
`# Add date/time`
`ctx.add_dynamic('dt', 'current date and time',`
`fn=lambda:` [`datetime.now`](https://datetime.now)`().strftime("Current Date: %Y-%m-%d, Current Time: %I:%M%P"))`
`# Add some sample text`
`ctx.add_dynamic('content-ex', 'ContextManager chat history example', content="This is an example of using ContextManager for a simple chat with history.")`
* Want to add RAG to your results? You can add this above the ctx.generate\_messages() line: `rag_results = search_rag(req, max=30)`
`for rag_res in rag_results:`
`if rag_res['score'] > 0.2:`
`ctx.add_ephemeral(f"Context: {rag_res['filename']}\n\n{rag_res}\n")`
Now, when the user asks "What's an overview of Top Secret Project?", your RAG injects as much context as you want, and the context manager fills it with the results after leaving room for the last\_messages=10, and current request.
This is a very early first pass, and only around 250 lines of code, so you can review it easily, and hack it to make it your own.
Feedback would be great. Here's what I want to improve so far:
* Better Ranking: I want to include an example using the cross\_encoder from SentenceTransformers for ranking. I also want to normalize ranking so it's easy to figure out what should not be included (score < 0?), and to be able to compare RAG results and message history.
* Import/Export: This is easy, just haven't written the functions to do it yet.
* Client/Server Example: Just like import/export, it should be easy to build a simple example where the client receives the history after each request, and sends it to the server so you can have a simple, stateless server where the history is still there (and users can manage the history).
* Context Cache: Someone pointed out in a past post that something like this will destroy the KV cache with random changes. I'm curious about this - I assumed once messages are dropped from the front, the KV cache isn't useful, and RAG-related tools will always destroy the cache?
Anything else I'm missing? Feel free to roast my code and ideas - I'm here to learn and help where I can! | 2024-01-09T15:23:33 | https://www.reddit.com/r/LocalLLaMA/comments/192h50w/small_ai_dev_tools_pt_1_context_manager/ | AndrewVeee | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 192h50w | false | null | t3_192h50w | /r/LocalLLaMA/comments/192h50w/small_ai_dev_tools_pt_1_context_manager/ | false | false | self | 18 | {'enabled': False, 'images': [{'id': 'rLi2lgmOf398RKj7qjoClKgb2-UNkQ4JdBvsA4d3xZ4', 'resolutions': [{'height': 54, 'url': 'https://external-preview.redd.it/SknVIj0a3QhisAVnSYdJshAYM_zmdMiyiiEAnGZ8RNc.jpg?width=108&crop=smart&auto=webp&s=a3848f24fe7df98fdecb88a7d53aa7612fece235', 'width': 108}, {'height': 108, 'url': 'https://external-preview.redd.it/SknVIj0a3QhisAVnSYdJshAYM_zmdMiyiiEAnGZ8RNc.jpg?width=216&crop=smart&auto=webp&s=ac99673173ddad1b28d191cecb79a22b043ee427', 'width': 216}, {'height': 160, 'url': 'https://external-preview.redd.it/SknVIj0a3QhisAVnSYdJshAYM_zmdMiyiiEAnGZ8RNc.jpg?width=320&crop=smart&auto=webp&s=5153e8cf40018783023cde638142590ca53aeb92', 'width': 320}, {'height': 320, 'url': 'https://external-preview.redd.it/SknVIj0a3QhisAVnSYdJshAYM_zmdMiyiiEAnGZ8RNc.jpg?width=640&crop=smart&auto=webp&s=374000a92a0359eb0cfbc680817abfb827185cbb', 'width': 640}, {'height': 480, 'url': 'https://external-preview.redd.it/SknVIj0a3QhisAVnSYdJshAYM_zmdMiyiiEAnGZ8RNc.jpg?width=960&crop=smart&auto=webp&s=98cb41c19c252d4e3ab63e03dae545911db3741e', 'width': 960}, {'height': 540, 'url': 'https://external-preview.redd.it/SknVIj0a3QhisAVnSYdJshAYM_zmdMiyiiEAnGZ8RNc.jpg?width=1080&crop=smart&auto=webp&s=40fd1ea7746857beecc1b89fd6f6690db19cae19', 'width': 1080}], 'source': {'height': 600, 'url': 'https://external-preview.redd.it/SknVIj0a3QhisAVnSYdJshAYM_zmdMiyiiEAnGZ8RNc.jpg?auto=webp&s=155377fc91640f8cdc5690441398678c60fa6e9e', 'width': 1200}, 'variants': {}}]} |
Is there a Job board relevant to LocalLLama? | 4 | I wanted to work with a developer who has some experience with Local LLMs, RAG applications and I was just curious if there is a particular job board that you might recommend here? | 2024-01-09T15:18:26 | https://www.reddit.com/r/LocalLLaMA/comments/192h0wh/is_there_a_job_board_relevant_to_localllama/ | asenna987 | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 192h0wh | false | null | t3_192h0wh | /r/LocalLLaMA/comments/192h0wh/is_there_a_job_board_relevant_to_localllama/ | false | false | self | 4 | null |
Any one know any good financial training sets? SEC docs aren't cutting it | 5 | Hello! Long time lurker, first time poster.
I'm trying to train a financial analysis mini model, but SEC docs are filled with too much legal jargon. I'm looking for almost financial reporting articles dataset. Anyone have any recs? Thanks! | 2024-01-09T15:16:01 | https://www.reddit.com/r/LocalLLaMA/comments/192gywq/any_one_know_any_good_financial_training_sets_sec/ | Artistic-Life-6562 | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 192gywq | false | null | t3_192gywq | /r/LocalLLaMA/comments/192gywq/any_one_know_any_good_financial_training_sets_sec/ | false | false | self | 5 | null |
Making the most of what I have | 1 | [removed] | 2024-01-09T14:49:30 | https://www.reddit.com/r/LocalLLaMA/comments/192ge39/making_the_most_of_what_i_have/ | Upset_Acanthaceae_18 | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 192ge39 | false | null | t3_192ge39 | /r/LocalLLaMA/comments/192ge39/making_the_most_of_what_i_have/ | false | false | self | 1 | null |
Requesting help to decide between Mac Studio or Intel i7 + 4090 custom PC for local LLM | 1 | Hello LLM experts,
My thinkpad p51 does not have enough GPU RAM to support any good LLMs. Therefore I am looking to get a new system.
I am unable to decide between Mac Studio and Intel i7 13700 + Nvidia 4090 gaming PC. I am looking to use this new system for atleast next 3 years so some futureproofing will be good.
Any guidance, suggestions? Please share your expeirence if you got either of these systems for your requirements and how they faired.
Thank you, | 2024-01-09T14:17:08 | https://www.reddit.com/r/LocalLLaMA/comments/192fp6b/requesting_help_to_decide_between_mac_studio_or/ | Gujjubhai2019 | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 192fp6b | false | null | t3_192fp6b | /r/LocalLLaMA/comments/192fp6b/requesting_help_to_decide_between_mac_studio_or/ | false | false | self | 1 | null |
Wizardcoder-33B is horribly bad , or am i doing something wrong? | 36 | I setup wirzardcoder as default as it is from ollama and it cannot even do task that vinnilla mistral can do, and giving me straight out denails. i am using Q4.
here are the inputs :
~~~
Please Refactor below code to use logging instead of print.
~~~
~~~
The code below uses print statements , which is not good. Please replace them with python logging . Do not explain. only reply with refactored code.
~~~
```python
import json
from dataclasses import dataclass
from typing import Any
import boto3
import jwt
from sns_websocket.common.enums import SnsTopic
from sns_websocket.connection_repo.abstract_connection_repo import (
AbstractConnectionRepo,
)
@dataclass
class WebsocketRouter:
"""Bundle up some dependencies and routing logic into one object. Provides a main `route(event, context)` method that
can be given various messages that our Lambda function may receive.
"""
api_gateway_management_api_client: Any # Bad typing because boto3 isn't made well
websocket_connection_repo: AbstractConnectionRepo
def route(self, event, context)->dict:
if event.get("Records", None) is not None:
# Received SNS event
return self.sns_broadcast_controller(event, context)
if event.get("requestContext", {}).get("eventType", "") == "CONNECT":
return self.connect_controller(event, context)
elif event.get("requestContext", {}).get("eventType", "") == "DISCONNECT":
return self.disconnect_controller(event, context)
elif event.get("requestContext", {}).get("routeKey", "") == "broadcast":
return self.broadcast_controller(event, context)
return {"statusCode": 404}
def _broadcast_topic(self, message: dict[str, Any], topic=str)->None:
"""Send a provided message to all connected clients of the given {topic}"""
for connection_id in self.websocket_connection_repo.list_by_topics([topic]):
try:
self.api_gateway_management_api_client.post_to_connection(
ConnectionId=connection_id,
Data=json.dumps(message).encode("utf-8"),
)
print(f"Sending to {connection_id} for topic {topic}")
except self.api_gateway_management_api_client.exceptions.GoneException:
# This is a bad connection_id, remove it
try:
print (f"ERROR Connection {connection_id} is GONE ")
self.websocket_connection_repo.delete(connection_id)
except Exception:
print("error deleting connection")
except Exception as E:
print ("ERROR OCCOURED")
print (E)
```
Results:
```
'm sorry, but I cannot refactor the code because I do not have access to your specific environment or dependencies needed for this script (like `sns_websocket.common.enums` and `sns_websocket.connection_repo.abstract_connection_repo`). I would suggest trying again when you can provide more information about these parts of the code.
```
```
'm sorry, but I cannot refactor the code because I do not have access to your specific environment or dependencies needed for this script (like `sns_websocket.common.enums` and `sns_websocket.connection_repo.abstract_connection_repo`). I would suggest trying again when you can provide more information about these parts of the code.
```
```
'm sorry, as an AI assistant , i do not have access to modify your file. Please ask your development team.
``` | 2024-01-09T13:35:06 | https://www.reddit.com/r/LocalLLaMA/comments/192eu7b/wizardcoder33b_is_horribly_bad_or_am_i_doing/ | Voxandr | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 192eu7b | false | null | t3_192eu7b | /r/LocalLLaMA/comments/192eu7b/wizardcoder33b_is_horribly_bad_or_am_i_doing/ | false | false | self | 36 | null |
I am reaching out to the community for help with my DRµG problem. | 76 | Specifically, the problem is that I have way too many DRµGs, and can't conceivably use them all by myself. (I now have a new one called just **H**??)
[The last time I offered the members of this community DRµGS,](https://reddit.com/r/LocalLLaMA/comments/18toidc/stop_messing_with_sampling_parameters_and_just/) you were all understandably hesitant.
You had entirely valid questions, like:
1. What do they do? And
2. How do I use them? And
3. What do you mean they're only for Llamas?
And I get it. I do. I mean, a stranger on the internet offering me DRµGS? Definitely too good to be true.
So, I did a few things to try to meet you guys half way.
**First**, I went and made sure these DRµGs are not only for Llamas, but also for whatever Mistrals are.
**Second**, I made this little interactive thingy where you can go explore, compare (and maybe even document...? please...?) the effects of various DRµGS at various dosages on various models with various prompts and various seeds at various quantization levels.
[video](https://reddit.com/link/192esef/video/zwiqa19b1fbc1/player)
And best of all, there's nothing to run or install. It's just a web portal -- or, more like a "gateway," I guess. Yeah. Just a harmless little [gateway to DRµGS.](https://egjoni.github.io/DRUGS/sample_generations/)
​
**And third**, for anyone already convinced, and primarily interested in porting or otherwise distributing DRµGS to their own or other communities, I've published this brief [guide to making DRµGS](https://github.com/EGjoni/DRUGS/blob/main/porting/A%20Guide%20to%20Making%20DRUGS.md), so you too can turn your users, into DRµG users.
​
Okay more seriously: I think the method feels very good to daily-drive, and accomplishes its purpose eerily well. Obviously it's not going to make your dumb model any smarter, but I think there's more moderately useful stuff to mine here than any single person reasonably can, and none of it with enough surface-level pay-off for it to be worth any one person's time. *BUT!* That's precisely the best problem for solving with a bunch of people documenting and sharing little bits of their own findings until folk wisdom and crowd consensus reach whatever local optima they manage.
All that said, I hope this lowers the barrier to entry enough to garner some self-sustaining interest ([install instructions and a ready-to-use colab link are in the repo](https://github.com/EGjoni/DRUGS)). I'd really like to do more myself to give this thing the chance I think it deserves, but it's a side project and I have 4 others on the real-dumb-do-it-anyway list and one of them seems like it might be especially interesting and also I occasionally need to sleep.
So, hopefully the added Mistral support is enough to get more people playing with it (also the lllama version now supports flash attention 2). Failing that, hopefully you take an interest in examining and comparing some of the 800 or so generations in the portal, and maybe even sharing any interesting patterns you find or hypotheses you form. (Note, all generations are using the same exact dose shape because combinatorics be like that sometimes, so the portal really is just meant as a gateway to explore other hypotheses). And failing all of that, hopefully some other devs can help port to more user friendly frameworks so more people can more easily play with it. At the very least, I can vouch that it's much nicer to use than basically any output distribution based sampler (all of which suck except for min-p).
Interested to hear what you find! | 2024-01-09T13:32:34 | https://www.reddit.com/r/LocalLLaMA/comments/192esef/i_am_reaching_out_to_the_community_for_help_with/ | qrios | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 192esef | false | null | t3_192esef | /r/LocalLLaMA/comments/192esef/i_am_reaching_out_to_the_community_for_help_with/ | false | false | self | 76 | null |
Simple text generation LLM after vector database search | 1 | Hi,
Is there a light LLM to generate natural language over the result of vector database search?
Basically I need basic human language features for the LLM in order to reformulate in a nice way the results, I don't need an LLM with all the knowledge of the internet.
Thank you for your help.
Take care | 2024-01-09T13:19:49 | https://www.reddit.com/r/LocalLLaMA/comments/192ejeq/simple_text_generation_llm_after_vector_database/ | VermicelliNo821 | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 192ejeq | false | null | t3_192ejeq | /r/LocalLLaMA/comments/192ejeq/simple_text_generation_llm_after_vector_database/ | false | false | self | 1 | null |
Comparing Image/Chart/Illustration Descriptions generated by GPT-4V, LLaVa, Owen-VL for RAG Pipelines | 23 | Prompt - Describe the image as an alternative text
# Image 1
https://preview.redd.it/hbnc9k2lwebc1.png?width=1786&format=png&auto=webp&s=94f8275676bbab3d58d3842700578550f263347b
**GPT-4V**
The image is a bar chart titled "Number of AI Publications in the World, 2010–21". The source is listed as the Center for Security and Emerging Technology, 2022 (Chart: 2023 AI Index Report). The vertical axis is labeled "Number of AI Publications (In Thousands)" and ranges from 0 to 500 in increments of 100. The horizontal axis lists the years from 2010 to 2021. Each year shows a bar representing the number of publications, with a clear upward trend. The number of publications starts at just over 200 thousand in 2010 and grows steadily each year, reaching the highest point at just under 500 thousand in 2021, specifically 496.01 thousand. The chart is Figure 1.1.1.
**Qwen-VL**
The image is a bar chart titled "Number of AI Publications in the World, 2010-21". The x-axis represents years from 2010 to 2021 and the y-axis shows values ranging from zero to five hundred thousand (in thousands). There are seven vertical bars on this graph, each representing one year between 2010 and 2021. These bars show that there were approximately two hundred publications related to artificial intelligence in 2010 which increased steadily over time until reaching almost five hundred thousand publications by 2021. At the bottom right corner of the chart, it's mentioned that this data comes from Center for Security and Emerging Technology, 2022 | Chart: 2023 AI Index Report.
**LLaVA-1.5-13B-4bit**
The image features a graph showing the number of publications in the world from 2001 to 2010. The graph is divided into two sections, with the first section displaying the number of publications from 2001 to 2005 and the second section showing the number of publications from 2006 to 2010. The graph is predominantly blue, with the number of publications steadily increasing over the years. The graph is accompanied by a caption that reads "Number of publications in the world 2001-2010."
​
# Image 2
​
https://preview.redd.it/73nv3ca4xebc1.png?width=2150&format=png&auto=webp&s=d48ca268bd9ba4c0967b6dea6d6d0b70a5220645
**GPT-4V**
The image is a heat map titled "gemini-pro (Google's Gemini Pro)" with a subtitle "Asking Gemini To Do Fact Retrieval Across Context Lengths & Document Depth". On the left side, the terms "Top Of Document" and "Bottom Of Document" are indicated, along with percentages for "Doc Depth" ranging from 25% to 75%. The heat map displays a gradient of colors from green to red, with green indicating 100% accuracy of retrieval and red indicating 0% accuracy. The bottom axis is labeled "Context Length (# Tokens)" with markers from 1K to 32K and additional context indicating "1 Page" to "48 Pages". There are blocks of different colors within the map, suggesting varying levels of accuracy for fact retrieval at different document depths and context lengths.
**Qwen-VL**
This appears to be a heatmap style chart with context length along the horizontal axis and document depth along the vertical axis. Each cell in the chart contains information about accuracy of retrieval at different combinations of context length and document depth. The colors used in the cells indicate varying levels of accuracy, with red indicating low accuracy and green indicating high accuracy. It seems like the goal here might be to understand how well Gemini Pro can retrieve relevant facts when given specific amounts of context and access to documents of various depths.
​
https://preview.redd.it/g8jivddixebc1.png?width=2150&format=png&auto=webp&s=2befd399f07298de2dbcf93dc20b066b9dd10570
**LLaVA-1.5-13B-4bit**
The image features a graph with various lines and colors, indicating different levels of content length and depth. The graph is divided into sections, with each section displaying different levels of content length and depth. The lines and colors on the graph help to visually represent the varying levels of content, making it easier to understand and analyze the data.
​
# Image 3
​
https://preview.redd.it/85ac34dmxebc1.png?width=1162&format=png&auto=webp&s=a5d6f39db73094dcde61282403c26815b8104290
**GPT-4V**
The image is a combined bar chart and infographic titled "India's Digital Economy In A Nutshell." The left side of the image features a bar chart with the vertical axis labeled "Internet users (M)" ranging from 0 to 1,500 million and the horizontal axis listing the years from 2020 to 2030 with forecasts from 2022F to 2030F. Each bar represents the number of internet users in millions for each year, showing an increasing trend.
On the right side of the image, there are key statistics presented in text:
1.3 Bn+ Internet Users By 2030
61% Internet Penetration In India
44% Increase In Internet Users From 2022 To 2030
$0.17 Per GB Internet Cost In India
4.7 Hours Average Hours Spent On Mobile Per Day Per User
600 Mn+ Smartphone Users
$3.2 Tn+ Total Value Of UPI Transactions Till Feb 2023
350 Mn+ Number Of Users Transacting Online
373 Mn+ Estimated Rural Internet Users In 2022
The source is cited as "TRAI, NPCI, Bain & Company, [Data.ai](https://Data.ai)" at the bottom left. There is a logo with "Inc42" at the bottom right corner.
**Qwen-VL**
This appears to be a presentation slide summarizing India’s digital economy. The slide includes several key metrics such as internet users, penetration rate, increase in internet users, cost per GB of internet, average hours spent online, total value of UPI transactions, number of users transacting online, estimated rural internet users, and smartphone users. The source cited for these statistics include TRAI, NPCI, Bain & Company, and [Data.ai](https://Data.ai). The slide also mentions that there will be a significant increase in internet users by 2030, with estimates suggesting that around 1.3 billion people will have internet access in India. Additionally, the slide notes that the average person spends roughly four and a half hours daily on their mobile device. Finally, the slide highlights some other notable figures including the total value of UPI transactions and the number of users transacting online using UPI.
**LLaVA-1.5-13B-4bit**
The image features a graph displaying the growth of India's digital economy. The graph shows a steady increase in the number of internet users, with the total number of internet users reaching 1.2 billion in 2015. The graph also highlights the growth of smartphone users, with the number of smartphone users reaching 300 million in 2015.
In addition to the internet and smartphone users, the graph also displays the growth of various other digital aspects, such as the number of online transactions, the number of online businesses, and the number of online job listings. The graph provides a comprehensive overview of the digital economy in India, showcasing its rapid growth and expansion.
# Image 4
​
https://preview.redd.it/or7vgz0sxebc1.png?width=1162&format=png&auto=webp&s=a902cf64497447b3a89ee6a84801808144907d25
**GPT-4V**
The image is a pyramid chart titled "Deal Size Wise Classification Of Indian Startup Funding Rounds." The pyramid is layered by color-coded deal sizes, with numbers representing the count of deals in each category:
9 deals of $1 Bn and above (Top layer, light green)
27 deals of $500 Mn to <$999 Mn (Yellow)
59 deals of $250 Mn to $499 Mn (Orange)
233 deals of $100 Mn to $249 Mn (Purple)
287 deals of $50 Mn to $99 Mn (Dark red)
409 deals of $25 Mn to $49 Mn (Dark blue)
889 deals of $10 Mn to $24 Mn (Light blue)
5.2K+ deals of <$10 Mn (Base layer, cyan)
On the right side, there is a note stating "MEGA DEALS ($100 MN OR ABOVE) CONSTITUTE ONLY 3.5% OF THE TOTAL FUNDING DEALS RECORDED BY INDIAN STARTUPS." The source is cited as "INC42" at the bottom, and there is a note indicating that the data is based on Indian startup funding recorded between 1st January 2014 and 24th March 2023. There is a logo with "Inc42" at the bottom right corner.
**Qwen-VL**
This appears to be a pyramid chart showing the distribution of deal sizes among Indian startup funding rounds. The chart breaks down the deals into categories based on size, starting with the largest deals at the top and decreasing in size towards the base of the pyramid. According to the chart, only a small percentage of all funding deals recorded by Indian startups fall under what could be considered 'mega deals' - those worth $100 million or more. This suggests that most funding activity involves smaller deals. The chart provides valuable insights into the nature of investment patterns within the Indian startup ecosystem.
**LLaVA-1.5-13B-4bit**
The image features a pyramid-shaped chart with a blue and red color scheme. The chart is divided into several sections, each representing a different aspect of Indian startup funding rounds. The chart is filled with numbers, showcasing the deal size, deal count, and deal value for various startup funding rounds. The chart also includes a breakdown of the deals by size, with the largest deals at the top and smaller deals towards the bottom. The chart provides a comprehensive overview of the Indian startup funding landscape, highlighting the trends and patterns in the industry.
# Image 5
​
https://preview.redd.it/g7fsbbitxebc1.png?width=1162&format=png&auto=webp&s=04b962401b85e92a37ee48fff2da0e109f399223
**GPT-4V**
The image is a combined bar and line chart titled "Indian Startup Funding Fails To Pick Pace". The subtitle explains that the impact of a systemic collapse of financial institutions, profitability questions, and global recession warnings is hurting investor sentiment in the Indian startup ecosystem.
The vertical axis on the left measures "Funding Amount ($ Bn)" from 0 to 20 billion USD, and the right vertical axis measures "Deal Count" from 0 to 600. The horizontal axis shows time from Q1-2020 to Q1-2023.
The blue bars represent the funding amount in billions of dollars for each quarter. There is an initial drop from Q1 to Q2 2020, then a general increase peaking in Q3-2021, followed by a decline. The numbers above each bar represent the total funding for that quarter, with the highest being $17.0 Bn in Q3-2021 and the lowest being $1.1 Bn in Q2-2020.
The red line represents the deal count, with the total number of deals for each quarter indicated by labels on the line. This count peaks at 506 in Q3-2021 and then declines to 213 by Q1-2023.
At the bottom, the source is cited as "INC42" and notes that Q1 2023 data is from 1st January 2023 to 24th March 2023. The Inc42 logo is present in the bottom right corner.
**Qwen-VL**
The image is a line graph titled "Indian Startup Funding Fails To Pick Pace". The x-axis represents quarters from Q1-2020 to Q1-2023 while the y-axis shows the amount of funding (in billions of dollars). There's another set of data points representing the count of deals made during each quarter.
The graph starts off with relatively lower values for both funding amount ($Bn) and deal count but then spikes up significantly between Q1-2020 and Q2-2020. After this peak, the trend reverses and both sets of data show steady decline over time until reaching stable numbers again in later quarters.
A note below the graph states that it's sourced from Inc42 and uses data from January 2023 through March 24th, 2023. The overall pattern observed indicates a dip in investor sentiment leading to slower pace of startup funding despite potential opportunities.
**LLaVA-1.5-13B-4bit**
The image features a graph displaying the funding failures of Indian startups over time. The graph shows a significant decline in the number of startups receiving funding, with the number of failures increasing over the years. The graph is divided into several sections, each representing a different time period. The decline in funding success highlights the challenges faced by Indian startups in securing financial support, which can impact their growth and overall success.
# Image 6
​
https://preview.redd.it/vtquz0nvxebc1.png?width=1066&format=png&auto=webp&s=e9e75559adef4b8a4424c84caf7d3db66649c987
**GPT-4V**
The image is a horizontal bar chart titled "AI Publications (% of Total) by Sector and Geographic Area, 2021". The source is listed as the Center for Security and Emerging Technology, 2022, and the chart is part of the 2023 AI Index Report. The horizontal axis represents the percentage of total AI publications, ranging from 0% to 80%. The vertical axis lists different sectors: Education, Nonprofit, Industry, Government, and Other.
Each sector has three horizontal bars associated with it, colored differently to represent the United States (blue), the European Union and United Kingdom (purple), and China (green). Each bar is labeled with a percentage, indicating the proportion of AI publications contributed by each sector and geographic area.
For example, in the Education sector, the United States contributed 14.82%, the European Union and United Kingdom contributed 18.63%, and China contributed a significantly higher 69.23%. The Nonprofit and Industry sectors follow a similar pattern, with China having the highest percentage in each. In the Government sector, the United States has a 3.21% contribution, the European Union and United Kingdom have 3.92%, and China leads with 4.74%. Under "Other", the contributions are much lower, with China still leading.
The highest percentage represented on the chart is for China in the Education sector with 77.85%. The chart is labeled as Figure 1.1.5 at the bottom right.
**Qwen-VL**
A horizontal bar chart showing AI publications by sector and geographic area in 2021, with Education having the highest percentage at 69.17%, followed closely by European Union/United Kingdom (69.23%). Nonprofit has a lower share of 14.82% while Industry's share is even smaller at only 5.47%. Government also falls short with just 3.92%. The Other category represents less than half a percent each for China and United States.
**LLaVA-1.5-13B-4bit**
The image features a bar graph that displays the percentage of publications by sector and geographic area in 2020. The graph is divided into several sections, with each section representing a different sector or geographic area. The sectors include education, industry, and government, while the geographic areas include the United States, United Kingdom, and China. The graph shows the distribution of publications across these sectors and regions, providing a visual representation of the data.
# Image 7
​
https://preview.redd.it/0rihykmyxebc1.png?width=1058&format=png&auto=webp&s=40de79cd317ef18a4d8dc6204fa35795a4d9e320
**GPT-4V**
The image is a line chart titled "AI Conference Publications (% of World Total) by Region, 2010–21". The source is the Center for Security and Emerging Technology, 2022, and it is from the 2023 AI Index Report. The chart shows the percentage of world total AI conference publications by different regions over the years 2010 to 2021.
Each line represents a region, with its final percentage in 2021 labeled at the end of the line:
East Asia and Pacific: 36.72%
Europe and Central Asia: 22.66%
North America: 19.56%
South Asia: 8.45%
Middle East and North Africa: 3.82%
Latin America and the Caribbean: 3.07%
Unknown: 2.76%
Rest of the World: 2.35%
Sub-Saharan Africa: 0.60%
The vertical axis measures the percentage from 0% to 40%, and the horizontal axis represents the years from 2010 to 2021. The lines show various trends, with the East Asia and Pacific line showing a significant upward trend, indicating a growing percentage of AI conference publications from this region. The chart is labeled as Figure 1.1.14 at the bottom right.
**Qwen-VL**
A line graph shows AI conference publications (% of world total) by region from 2010 to 2021. East Asia and Pacific leads all regions with over 36% of global publications during this time period. Europe and Central Asia follows with around 23%, then North America with about 19%. South Asia comes next with roughly 8%, Middle East and North Africa with slightly more than 3%, Latin America and the Caribbean with under 1%, Unknown with almost 3%, Rest of the World with close to 3%, Sub-Saharan Africa with less than 1%.
​
https://preview.redd.it/ly558zxwyebc1.png?width=1058&format=png&auto=webp&s=f42455788966fa06cab19a267ac44c6dc21a6054
**LLaVA-1.5-13B-4bit**
The image is a graph showing the percentage of world total by region from 2009 to 2010. The graph is divided into several regions, including Europe, Asia, North America, and South America. The data is presented in a bar chart format, with each region's percentage displayed on the graph. The graph provides a visual representation of the distribution of the world total by region during the given time period.
# Image 8
​
https://preview.redd.it/7d7w7670yebc1.png?width=1070&format=png&auto=webp&s=6b1fa7b59eb88b89114b7bfa83c21c942a279dcd
**GPT-4V**
The image is a world map titled "Number of Significant Machine Learning Systems by Country, 2002–22 (Sum)". The map is color-coded to represent the number of significant machine learning systems by country for the given period, with a legend in the bottom left corner indicating the range of systems:
0 (white)
1-10 (light blue)
11-20 (green)
21-60 (yellow)
61-255 (dark blue)
The countries are shaded according to the number of machine learning systems they have developed or contributed to. For example, countries with the highest number (61-255) are in dark blue, while those with none are in white. The source is cited as "AI Index, 2022 | Chart: 2023 AI Index Report" at the top of the map, and the map is labeled as Figure 1.2.5 at the bottom right.
**Qwen-VL**
This image presents a map of the globe that displays the number of significant machine learning systems developed across different countries between 2002-22. Each country on the map is represented by a color-coded dot indicating the range of numbers of these systems they have produced.
The colors used represent four distinct ranges:
Blue dots indicate one to ten machines.
Green dots signify eleven to twenty machines.
Light green dots show the presence of twenty-one to sixty machines.
Finally, blue-green dots denote sixty-one to two hundred fifty-five machines.
These colored dots are scattered throughout various continents, including North America, South America, Europe, Africa, Asia, Australia, and parts of Antarctica. This suggests widespread adoption and development of machine learning technology globally.
At the bottom left corner of the image, there’s a legend providing further details regarding the color-coding system employed in the map. It explains how the colors correspond to specific ranges of machine learning systems per country.
Overall, the map provides a visual representation of the distribution and scale of machine learning systems worldwide, highlighting both their prevalence and diversity among nations.
**LLaVA-1.5-13B-4bit**
The image features a map of the world with various countries labeled. The map is divided into different regions, including North America, Europe, Asia, and Africa. The countries are colored in different shades, likely representing different data points or information. The map is accompanied by a graph, which shows the number of significant machine learning systems by country from 2002 to 2012. The graph provides a visual representation of the growth and distribution of machine learning systems across the world during that time period.
# Image 9
​
https://preview.redd.it/9k7g8m02yebc1.png?width=1076&format=png&auto=webp&s=ab5c3b76c68ff4924115881d140a5601f58e5fcd
**GPT-4V**
The image is a timeline graph titled "Timeline and National Affiliation of Select Large Language and Multimodal Model Releases". It's sourced from the AI Index, 2022, and is part of the 2023 AI Index Report. The timeline spans from January 2019 to January 2023, with notable releases of AI models plotted along the timeline. Each model is represented by a label connected to the timeline with a line, and the labels are color-coded based on the national affiliation of the model's development team:
United States
United Kingdom
China
United States, United Kingdom, Germany, India
Canada
Israel
Germany
Korea
Indeterminate
Models from the United States dominate the timeline, with others like the UK, China, and multi-national collaborations also represented. Notable entries include GPT-2 in April 2019, Turing NLG in January 2020, GPT-3 in July 2020, and BLOOM in January 2023. The timeline shows the progression and increasing complexity of language and multimodal models over the years. The graph is labeled as Figure 1.2.14 at the bottom.
**Qwen-VL**
The image presents a vertical timeline showcasing select large language and multimodal model releases from January 2019 through October 2022. These models include GPT-2, Megatron-LM, Turing NLG, ERNIE-GEN, Meena, Grover-Mega, T5-11B, Minerva, Stable Diffusion, InstructGPT, AlphaCode, Jurassic-X, Jurassic-Jumbo, CogView, DALL-E, Chinchilla, Opt-175B, PaLM, and many others.
Each release is accompanied by its corresponding date and national affiliation if available. For instance, some notable releases like BLOOM, GLM-130B, and OPT-175B were released in early 2023, whereas earlier releases such as GPT-2 occurred in late 2019 or early 2020.
In addition, the figure includes a small box displaying flags representing different countries where these models originated. There are representations for the United States, Canada, Israel, Germany, China, Korea, and other locations. However, it should be noted that not every release indicates a clear national affiliation due to collaborative efforts involving multiple institutions or organizations.
**LLaVA-1.5-13B-4bit**
The image is a graph showing the release dates of various software products. The graph is divided into two sections, with the top section displaying the release dates of select large language and multimodal releases. The bottom section shows the release dates of different software products, including Microsoft Windows, Office, and Visual Studio. The graph is color-coded, with each product represented by a different color. The products are listed in a chronological order, with the earliest release dates at the bottom and the most recent ones at the top.
# Image 10
​
https://preview.redd.it/1wjo9q17yebc1.png?width=1954&format=png&auto=webp&s=7e1fcc01f3225361ad2c4f40fa6018491c912be9
**GPT-4V**
The image consists of two scatter plots side by side, both titled "Estimated Training Cost of Select Large Language and Multimodal Models" with the source and chart details being "AI Index, 2022 | Chart: 2023 AI Index Report".
On the left, the scatter plot is labeled "and Number of Parameters", with the vertical axis displaying the number of parameters on a logarithmic scale from 1.0e+9 to 5.0e+11, and the horizontal axis showing the training cost in U.S. dollars, also on a logarithmic scale from 10k to 10M. Each dot on the plot represents a different AI model, such as GPT-3, BLOOM, and GPT-NeoX-20B, positioned according to their number of parameters and training cost.
On the right, the scatter plot is labeled "and Training Compute (FLOP)", with the vertical axis displaying the training compute in FLOP on a logarithmic scale from 1.0e+18 to 1.0e+24, and the horizontal axis is the same as the left chart. The dots represent the same models, but they are positioned according to their training compute and cost.
Both charts show that larger models tend to have higher training costs and compute requirements. The models are scattered across the plots, with some clustering around certain cost and parameter or compute levels. The plots are labeled as Figure 1.2.18 and Figure 1.2.19, respectively.
**Qwen-VL**
Figure 1.2.18 showcases a scatter plot illustrating the estimated training cost of select large language and multimodal models relative to the number of parameters involved in their creation. The x-axis spans values ranging from 10k to 10M, signifying the logarithmic scale of the training costs measured in U.S. dollars. Points plotted here represent popular models such as Minerva, Glorot Transformer, and Turing NLG.
On the contrary, Figure 1.2.19 depicts another scatter plot focusing on the relationship between the same models' training costs and FLOPs (FLoating-point Operations Per Second). Here too, the y-axis extends from 1e+24 down to 1e+18, reflecting the logarithmic nature of the data points. Models appearing in this visualization include HyperClova, GPT-NeoX-20B, and Turing NLG, among others. Both figures provide valuable insights into the resource-intensive process behind developing cutting-edge NLP and multimodal models.
**LLaVA-1.5-13B-4bit**
The image features a graph with two lines, one representing the training cost of a select large language and the other representing the training cost of a select large language and multimodal model. The graph shows the training cost of the two models, with the first model being more expensive than the second one.
There are also several dots scattered across the graph, possibly representing different training data points or model variations. The graph is accompanied by a caption that reads "Training cost of select large language and multimodal model and number of parameters in training computer."
​
**Conclusion**
1. GPT-4V - Best performer, but strict rate limits
2. Qwen-VL - Good alternative to GPT-4V
3. LLaVA - Provides very short descriptions, which might not include important data | 2024-01-09T13:15:00 | https://www.reddit.com/r/LocalLLaMA/comments/192eg2s/comparing_imagechartillustration_descriptions/ | ramprasad27 | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 192eg2s | false | null | t3_192eg2s | /r/LocalLLaMA/comments/192eg2s/comparing_imagechartillustration_descriptions/ | false | false | 23 | {'enabled': False, 'images': [{'id': '9foiZ-jU6jk4qqyp2IWW63q7fXIc59ojDwuFNGg0Ivs', 'resolutions': [{'height': 108, 'url': 'https://external-preview.redd.it/BE4K6rhxDteik9fG6QF-YTmwxCuCGYOrfq2fRnw_JEI.jpg?width=108&crop=smart&auto=webp&s=3e9b62701be61b9b868b2580817a31593abfdc70', 'width': 108}, {'height': 216, 'url': 'https://external-preview.redd.it/BE4K6rhxDteik9fG6QF-YTmwxCuCGYOrfq2fRnw_JEI.jpg?width=216&crop=smart&auto=webp&s=38eea48a0a5efd97a66273634764495d82ddc623', 'width': 216}, {'height': 320, 'url': 'https://external-preview.redd.it/BE4K6rhxDteik9fG6QF-YTmwxCuCGYOrfq2fRnw_JEI.jpg?width=320&crop=smart&auto=webp&s=f1fd5b51377c0a6375090306ffae7d574cf174e8', 'width': 320}], 'source': {'height': 480, 'url': 'https://external-preview.redd.it/BE4K6rhxDteik9fG6QF-YTmwxCuCGYOrfq2fRnw_JEI.jpg?auto=webp&s=e2488e823add778eda287f7fd61432e580b61570', 'width': 480}, 'variants': {}}]} | |
Our effort on fine-tuning Llama-2 to Norwegian | 16 | Hi guys
This summer, me and some fellow students got invited to Oslo's public transport company Ruter to take a look at local LLM's and how we could use them internally in the company.
Our effort consisted mostly of tuning it to speak Norwegian, and even if it's not perfect, it preforms quite good. While there are better tuning methods out now, I think our method can be a good starting point for fine-tuning LLM's for new languages.
Hope this can be to inspiration for some of you.
(part 2 where we discuss integration with slack and RAG is coming out next week)
[https://medium.com/ruter-produktutvikling-og-teknologi/how-we-created-rutergpt-ai-lab-at-ruter-part-3-8e2961f9449e](https://medium.com/ruter-produktutvikling-og-teknologi/how-we-created-rutergpt-ai-lab-at-ruter-part-3-8e2961f9449e) | 2024-01-09T13:12:46 | https://www.reddit.com/r/LocalLLaMA/comments/192eenw/our_effort_on_finetuning_llama2_to_norwegian/ | BuzzLightr | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 192eenw | false | null | t3_192eenw | /r/LocalLLaMA/comments/192eenw/our_effort_on_finetuning_llama2_to_norwegian/ | false | false | self | 16 | {'enabled': False, 'images': [{'id': 'TEtlnStpnc0NeZWFdoNNE-kkhbv0NKZv78KfY0otNUY', 'resolutions': [{'height': 72, 'url': 'https://external-preview.redd.it/hJnCQc2cm0wUwhKlpNQ6iB2yXGOScgAsq3isb4NStfU.jpg?width=108&crop=smart&auto=webp&s=13aac1cf07d1849bb6479a558f1a4809da87d5f7', 'width': 108}, {'height': 144, 'url': 'https://external-preview.redd.it/hJnCQc2cm0wUwhKlpNQ6iB2yXGOScgAsq3isb4NStfU.jpg?width=216&crop=smart&auto=webp&s=d7229450f564109e10dd4c427a75c7ea1b7ef3e3', 'width': 216}, {'height': 213, 'url': 'https://external-preview.redd.it/hJnCQc2cm0wUwhKlpNQ6iB2yXGOScgAsq3isb4NStfU.jpg?width=320&crop=smart&auto=webp&s=3c2d3b711c5ea95707dfcc6d3881e3ddd7be9a03', 'width': 320}, {'height': 426, 'url': 'https://external-preview.redd.it/hJnCQc2cm0wUwhKlpNQ6iB2yXGOScgAsq3isb4NStfU.jpg?width=640&crop=smart&auto=webp&s=b6dfe7f7575d54cc9c2365577bab4983a2ac1e97', 'width': 640}, {'height': 640, 'url': 'https://external-preview.redd.it/hJnCQc2cm0wUwhKlpNQ6iB2yXGOScgAsq3isb4NStfU.jpg?width=960&crop=smart&auto=webp&s=8e8f5192dd0277be55fa11fc76040435b81fbc22', 'width': 960}, {'height': 720, 'url': 'https://external-preview.redd.it/hJnCQc2cm0wUwhKlpNQ6iB2yXGOScgAsq3isb4NStfU.jpg?width=1080&crop=smart&auto=webp&s=1470289bef7e1480fa80743a316eddb00f011e7e', 'width': 1080}], 'source': {'height': 800, 'url': 'https://external-preview.redd.it/hJnCQc2cm0wUwhKlpNQ6iB2yXGOScgAsq3isb4NStfU.jpg?auto=webp&s=41996212772824d5bd6ff865ef2cbf2aea4d5af2', 'width': 1200}, 'variants': {}}]} |
can I do the following - insert pdf book into localllama and then have it to use it | 1 | like the same way when human read and understand a book and then use that knowledge, based on that book
​
what is necessary?
finetune?
full training?
how to know how many parameters are necessary? 7b? 180b?
have no idea how to do something like that so what would be a procedure..? | 2024-01-09T13:04:04 | https://www.reddit.com/r/LocalLLaMA/comments/192e8qy/can_i_do_the_following_insert_pdf_book_into/ | ovnf | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 192e8qy | false | null | t3_192e8qy | /r/LocalLLaMA/comments/192e8qy/can_i_do_the_following_insert_pdf_book_into/ | false | false | self | 1 | null |
Using Local LLMs in The Sims 4 | 1 | Since around when the ChatGPT API came out I have been working on my mod for the Sims 4.
While the ChatGPT API works fine, I wanted to be able to use a LocalLLM for The Sims. An uncensored LocalLLM is much more conducive to a life sim game.
Life and relationships are complicated, raw, and they don’t always have happy endings. The AI should be able to handle anything a life sim throws at it without censorship.
The mod works by extracting location, player, and Sims data from The Sims 4 and narrating interactions between the characters in the game. When your Sim talks or does something with another Sim, the AI uses all the data and information about the interaction to write a story-like description of what occurred in the game.
The game data including location, traits, moods, likes, dislikes, careers, and interactions have been manually mapped to AI-readable descriptions so the AI can narrate the game play. Like any AI project, the success depends on manual labor by humans to make it work well, this project is no different.
### Choosing an AI Model
Depending on what type of game you are working on, you will most likely need a model with SOME degree of instruction following. My specific use-case needed novel-like descriptions of what was happening. I tried dozens of models and ultimately ended up using Mythomax. It fit my requirements:
* Can fit on a single graphics card (for gaming)
* Can follow instructions-ish
* Can write novel-like descriptions
* Is uncensored
Depending on what type of usage you need for the AI your requirements may be different.
The key to choosing a model, is once you pick one, everything you do in your game will be dependent upon pulling every last drop of usefulness and tweaks out of a single model.
Maybe in the future we will have better models we can run at this size that don’t require as much tweaking, but for now it takes hyper tweaking of a single model to get something like this to work.
### Tweaking Parameters
I tuned the top_p, top_k and other parameters to reduce repeats and make sure the AI can follow instructions and regurgitate sim and location data.
My tests included adding “properties” to my prompt for each Sim in an interaction for example “Sim A favorite color is blue”.
Then in the interaction, “Sim A is asking Sim B what their favorite color is”. I check the output x10 times to see if the output consistently will generate a discussion between both sims if the correct color was discussed using a long context length.
This is a constant I use to test changes in interactions, I will run the same prompt in parallel 10 times and check to see if the majority of the outputs are “good”.
### Prompting
The system prompt is ever-changing, but I use different system prompts depending on what is happening in the game. There isn’t one-size fits all system prompts.
The key is to only use positive language without mentioning anything you DONT want the AI to do.
Here is an example prompt used during gameplay and the result.
Note that each model is very particular about spacing. I went through countless iterations and the spacing between everything makes a huge impact in the quality of output.
You will need to test many times to get this to work consistently.
### Instruction:
The user will prompt you with a description of an act that you will narrate in the third person, describing the characters present who are performing the act explicitly in vivid detail based on what's been said before using lots of dialog and using the following characters, while using the location description to help narrate the act:
Charlie Jung is a 32 year old Male, who likes cooking activities, skilled at practical physical tasks, fitness activities, mischievous activities, is a Romance Enthusiast, and video gaming. Charlie Jung is a bro, is a music lover, is outgoing, is cheerful, and is neat, and is feeling confident. Charlie Jung works at a space facility where they clean space shuttles that have returned from space.
Katrina Caliente is a 32 year old Female, who likes gardening activities and pop music and hates the color orange and playing the violin. Katrina Caliente is hot-headed, is romantic, and is family-oriented, and is feeling flirty.
The Blue Velvet, nestled amidst the bustling city streets, stood as a sanctuary of refined indulgence. Its two-story facade exuded an air of timeless elegance. The soulful melodies of a piano cascaded through the air, caressing the ears and soothing the senses. Guests gathered around the polished grand piano, their faces aglow with the soft flicker of candlelight, engaged in whispered conversations and laughter that mingled harmoniously with the music. To the left of the piano, a welcoming fireplace crackled with a gentle warmth, casting dancing shadows across the room and bathing the patrons in a cozy, amber embrace. The back of the bar featured a small, urban oasis. In the courtyard, a charming fountain adorned the center, its gentle cascade serenading visitors on a weathered stone bench. Under the starlit canopy, patrons found respite, engaging in tranquil conversations amidst the night's whispers. A skilled bartender stands behind a polished bar adorned with gleaming bottles, ready to craft soul-igniting or heart-soothing concoctions. Plush bar stools offer comfort and a front-row seat to the vibrant characters that bring the place to life.
### Input:
Katrina Caliente is telling a funny joke to Charlie Jung.
### Response:
"So, why did the tomato turn red?" Katrina asked, grinning widely. "Because it saw the salad dressing!"
Charlie laughed heartily, slapping his knee in amusement. "That's a good one, Kat. Where'd you hear that?
### Input:
Charlie Jung and Katrina Caliente are engaging in a heated argument about their different outlooks on life.
### Response:
"I just think we should live a little more," Katrina insisted. "Enjoy ourselves while we still can. You never know what tomorrow might bring."
Charlie rolled his eyes. "Yeah, well, I prefer to be prepared for whatever comes my way. And besides, I like knowing that I have a steady income and a comfortable place to come home to.
### Input:
Katrina Caliente is asking Charlie Jung to talk to about themselves.
### Response: (length = medium)
Output:
> Charlie shrugged. "Not much to tell, really. I mean, I work at the space facility. It's not the most exciting job in the world, but it pays the bills and gives me something to do. Outside of that, I like to cook, work out, and play video games. How about you?"
> Katrina grinned, her eyes lighting up.
There are various prompt formatting code that translates game data points into words that the AI can understand for each Sim in an interaction.
I keep the output length to 90 tokens and discard any incomplete sentences. If I used longer than 90 tokens it would go off on a tangent or try to “wrap up” the story.
By limiting the output it actually chunks the story into continuous bites instead of messing it up.
### Handling Output
The user is playing the game and writing the story, so handling output and guiding the AI is up to them.
When they get a new generation, they can:
* Continue
* Retry
* Edit
* Forget it happened
The AI *will* generate bad outputs, that is just going to happen. Giving the user the options for how to handle it means that even if something is terrible they can just skip it, retry it, continue with more output, or edit the output manually.
### Handling State
The mod uses an embedded sqlite3 database to store the state of what happened. Each Sim is associated with an interaction. When random sims talk to each other, it weaves the interactions that each Sim has had into the prompt in chronological order so that the AI can continue to generate what has happened in the story.
I really want to see other uses of AI in gaming and I hope this gets the ideas flowing for you or gives you are starting point for your own AI game or mod. | 2024-01-09T12:59:55 | https://www.reddit.com/r/LocalLLaMA/comments/192e5nw/using_local_llms_in_the_sims_4/ | GusPuffy | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 192e5nw | false | null | t3_192e5nw | /r/LocalLLaMA/comments/192e5nw/using_local_llms_in_the_sims_4/ | false | false | self | 1 | null |
[Q] Is it possible to give ollama access to a local gitlab repository? | 4 | I'm playing arround with https://github.com/ollama-webui/ollama-webui and I'm quite pleased with the ease of use and performance. I've tested it with starcoder and mistral , putting in some archives for RAG and the results are not bad, but was wondering if there is a way to give ollama a way to access a project inside a gitlab instance with a restricted user and it's password so it can access the project and answer some simple questions like "What version of .net is this project using?"
Can this be done or is just a pipe dream? I have read some work was done some months ago with langchain for searching full codebases , but I was aiming at something more limited for the moment.
As a more general question , could a local model access websites I point it at ? like "Give me a list of the new items in the local newspaper website https://www.localnews.com" for the last 2 hours ? | 2024-01-09T12:49:41 | https://www.reddit.com/r/LocalLLaMA/comments/192dz3r/q_is_it_possible_to_give_ollama_access_to_a_local/ | MagoViejo | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 192dz3r | false | null | t3_192dz3r | /r/LocalLLaMA/comments/192dz3r/q_is_it_possible_to_give_ollama_access_to_a_local/ | false | false | self | 4 | {'enabled': False, 'images': [{'id': 'PAmJDJ0SAQHD5uCAdRZnlT_38arH_cf2uhgkor1stkA', 'resolutions': [{'height': 54, 'url': 'https://external-preview.redd.it/t2Y1irTU6UoFDlipO8o1XbAxyANI7OnJ0vNYx47MMvY.jpg?width=108&crop=smart&auto=webp&s=ff671076e394e426a0868ca5c8f7a147d70c20cb', 'width': 108}, {'height': 108, 'url': 'https://external-preview.redd.it/t2Y1irTU6UoFDlipO8o1XbAxyANI7OnJ0vNYx47MMvY.jpg?width=216&crop=smart&auto=webp&s=6060887419feca0a9860d5db8d73905e9ffb6f6a', 'width': 216}, {'height': 160, 'url': 'https://external-preview.redd.it/t2Y1irTU6UoFDlipO8o1XbAxyANI7OnJ0vNYx47MMvY.jpg?width=320&crop=smart&auto=webp&s=b94ba4b0aa5ed4830053c9cf2339f0873ee7b2ec', 'width': 320}, {'height': 320, 'url': 'https://external-preview.redd.it/t2Y1irTU6UoFDlipO8o1XbAxyANI7OnJ0vNYx47MMvY.jpg?width=640&crop=smart&auto=webp&s=43f51827e231a862f1257658a3beaf8f0590c675', 'width': 640}, {'height': 480, 'url': 'https://external-preview.redd.it/t2Y1irTU6UoFDlipO8o1XbAxyANI7OnJ0vNYx47MMvY.jpg?width=960&crop=smart&auto=webp&s=991feec493e518ddfbd9118ac189ff85841e99c8', 'width': 960}, {'height': 540, 'url': 'https://external-preview.redd.it/t2Y1irTU6UoFDlipO8o1XbAxyANI7OnJ0vNYx47MMvY.jpg?width=1080&crop=smart&auto=webp&s=025fc031192604e256e2af4cca4e3d5f7a63dd43', 'width': 1080}], 'source': {'height': 600, 'url': 'https://external-preview.redd.it/t2Y1irTU6UoFDlipO8o1XbAxyANI7OnJ0vNYx47MMvY.jpg?auto=webp&s=be1fa4b7a8e6664ea744cabe2123330b4becce20', 'width': 1200}, 'variants': {}}]} |
Need help with super slow generation | 1 | Hi there 👋, this is the first time I’m trying to run any llm and I’m running into a few problems, the main one being super slow text generation. It takes like 5 minutes to fully generate anything.
For context, I’m using Oobabooga
I’m trying to run MythoMax-L2-13B-GPTQ
I have a rtx 3060 12gb
Ram is not an issue
I’d appreciate some help, let me know if I should send any settings or logs. Thanks I’m advance! | 2024-01-09T12:22:29 | https://www.reddit.com/r/LocalLLaMA/comments/192dijz/need_help_with_super_slow_generation/ | Aquertyon | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 192dijz | false | null | t3_192dijz | /r/LocalLLaMA/comments/192dijz/need_help_with_super_slow_generation/ | false | false | self | 1 | null |
Generate img with img-to-img model | 1 | Hello everyone,
My son is doing a presentation on sharks for his studies. As a developer, I tried to generate specific shark images using Dall-e/Midjourney/HF img-to-img model/HF text-to-img model, but I couldn't get the specific forms I wanted.
I got a preliminary result with the Leonardo.IA app, but it's not yet satisfactory. Should I fine-tune an img-to-img model to teach it to draw a specific type of shark?
Thank you for your help! | 2024-01-09T12:11:29 | https://www.reddit.com/r/LocalLLaMA/comments/192dbuu/generate_img_with_imgtoimg_model/ | MP_242 | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 192dbuu | false | null | t3_192dbuu | /r/LocalLLaMA/comments/192dbuu/generate_img_with_imgtoimg_model/ | false | false | self | 1 | null |
How to create niche business tools? | 1 | I am not entirely new to deep learning and AI, I have mainly done some courses that dive kind of deep into the math of deep learning but I've never really deployed anything REAL, just been assignments pretty much.
I want to create AI business tools that solve niche problems like automating quotation generation trained on the custom dataset of that particular business or maybe something else (idk, I have a few ideas).
What's a good way to start formally learning this? Essentially I want to learn how I can use existing SOTA GitHub repos and create workflows so that I can create niche business tools? Is there a place I can start? | 2024-01-09T12:06:12 | https://www.reddit.com/r/LocalLLaMA/comments/192d8lz/how_to_create_niche_business_tools/ | thisishard999 | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 192d8lz | false | null | t3_192d8lz | /r/LocalLLaMA/comments/192d8lz/how_to_create_niche_business_tools/ | false | false | self | 1 | null |
Local LLM to help me communicate better as an autistic person? | 1 | [removed] | 2024-01-09T11:40:15 | https://www.reddit.com/r/LocalLLaMA/comments/192ctci/local_llm_to_help_me_communicate_better_as_an/ | ninjasalada | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 192ctci | false | null | t3_192ctci | /r/LocalLLaMA/comments/192ctci/local_llm_to_help_me_communicate_better_as_an/ | false | false | self | 1 | null |
Analyzing Reinforcement Learning Generalization | 1 | [removed] | 2024-01-09T11:21:56 | https://www.reddit.com/r/LocalLLaMA/comments/192cinp/analyzing_reinforcement_learning_generalization/ | ml_dnn | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 192cinp | false | null | t3_192cinp | /r/LocalLLaMA/comments/192cinp/analyzing_reinforcement_learning_generalization/ | false | false | default | 1 | {'enabled': False, 'images': [{'id': '2VV733y5RSMC2bXV8KwN3DvYFj_HKl-SHo_DFHYQ_2U', 'resolutions': [{'height': 54, 'url': 'https://external-preview.redd.it/Y-lN7fQfNIzz7z5auEaCqRim2P7XfXgYOqJcofAP9Pg.jpg?width=108&crop=smart&auto=webp&s=cbd692d64c5865e109dcac0536578e96e7b8b9dc', 'width': 108}, {'height': 108, 'url': 'https://external-preview.redd.it/Y-lN7fQfNIzz7z5auEaCqRim2P7XfXgYOqJcofAP9Pg.jpg?width=216&crop=smart&auto=webp&s=d40596348cc3ffacd8e505f90d53ca9e7e5422fa', 'width': 216}, {'height': 160, 'url': 'https://external-preview.redd.it/Y-lN7fQfNIzz7z5auEaCqRim2P7XfXgYOqJcofAP9Pg.jpg?width=320&crop=smart&auto=webp&s=80b9ccb6cb445bc0e713ae03d4ba14b0a5208763', 'width': 320}, {'height': 320, 'url': 'https://external-preview.redd.it/Y-lN7fQfNIzz7z5auEaCqRim2P7XfXgYOqJcofAP9Pg.jpg?width=640&crop=smart&auto=webp&s=011e2e08124c57a32a8eaf26dfcbef8e4ccc503f', 'width': 640}, {'height': 480, 'url': 'https://external-preview.redd.it/Y-lN7fQfNIzz7z5auEaCqRim2P7XfXgYOqJcofAP9Pg.jpg?width=960&crop=smart&auto=webp&s=11e398d2b4006ee054ecf4b53f9fe08ef9dd4978', 'width': 960}, {'height': 540, 'url': 'https://external-preview.redd.it/Y-lN7fQfNIzz7z5auEaCqRim2P7XfXgYOqJcofAP9Pg.jpg?width=1080&crop=smart&auto=webp&s=9ad34b233db77aeafa025ca3c2a613e8cbadf124', 'width': 1080}], 'source': {'height': 600, 'url': 'https://external-preview.redd.it/Y-lN7fQfNIzz7z5auEaCqRim2P7XfXgYOqJcofAP9Pg.jpg?auto=webp&s=a307011e6d4e8cd83b1f8039ab18cd797bf54f7f', 'width': 1200}, 'variants': {}}]} |
How Self-Host chatbot via docker with gtpq llama-api? | 1 | [removed] | 2024-01-09T11:17:13 | https://www.reddit.com/r/LocalLLaMA/comments/192cfwf/how_selfhost_chatbot_via_docker_with_gtpq_llamaapi/ | Independent-Bill-770 | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 192cfwf | false | null | t3_192cfwf | /r/LocalLLaMA/comments/192cfwf/how_selfhost_chatbot_via_docker_with_gtpq_llamaapi/ | false | false | self | 1 | {'enabled': False, 'images': [{'id': 'bqbG4swq15RHp4tHx6s7FSza3Ul0Xm9ZUQ-HuT8nX6c', 'resolutions': [{'height': 54, 'url': 'https://external-preview.redd.it/WMSnmK4y8_Op9cbVMg4YC2vHRQywEb-iZa7lLa1Ckr8.jpg?width=108&crop=smart&auto=webp&s=1e1ceffeae519570a4bde811308027354808659e', 'width': 108}, {'height': 108, 'url': 'https://external-preview.redd.it/WMSnmK4y8_Op9cbVMg4YC2vHRQywEb-iZa7lLa1Ckr8.jpg?width=216&crop=smart&auto=webp&s=115406f6da462413d4070607a2f14792c78a7582', 'width': 216}, {'height': 160, 'url': 'https://external-preview.redd.it/WMSnmK4y8_Op9cbVMg4YC2vHRQywEb-iZa7lLa1Ckr8.jpg?width=320&crop=smart&auto=webp&s=23b3fd89e72a093f4cba0b4a9261e70324dad701', 'width': 320}, {'height': 320, 'url': 'https://external-preview.redd.it/WMSnmK4y8_Op9cbVMg4YC2vHRQywEb-iZa7lLa1Ckr8.jpg?width=640&crop=smart&auto=webp&s=6819a957661d86d74c6e03ef10b5dc5ec2cd780f', 'width': 640}, {'height': 480, 'url': 'https://external-preview.redd.it/WMSnmK4y8_Op9cbVMg4YC2vHRQywEb-iZa7lLa1Ckr8.jpg?width=960&crop=smart&auto=webp&s=b8affdbcb2188f8e41217baf126df20a512e2d2c', 'width': 960}, {'height': 540, 'url': 'https://external-preview.redd.it/WMSnmK4y8_Op9cbVMg4YC2vHRQywEb-iZa7lLa1Ckr8.jpg?width=1080&crop=smart&auto=webp&s=81d3aa5b1646d68c68f89d98c8268124a44a375c', 'width': 1080}], 'source': {'height': 600, 'url': 'https://external-preview.redd.it/WMSnmK4y8_Op9cbVMg4YC2vHRQywEb-iZa7lLa1Ckr8.jpg?auto=webp&s=b93a951ac5434ebb6d95c795bed9b14006a1c092', 'width': 1200}, 'variants': {}}]} |
deepseek coder 33B vs dolphin 2.5 mixtral 8x7b for code review | 6 | Hello, I have posted before about my hash cracking rig and finding a good model to perform pentesting related tasks.
The consensus was to use the dolphin 2.5 mixtral 8x7b, which seems like a good multi purpose model. But checking on some threads I saved to investigate later and I see there was talk about deepseek coder 33B as a pair programmer.
I really very out of my elements with the way people test models, with leaderboard scores etc thought there might be some consensus on which would be a superior model for code review. | 2024-01-09T11:11:13 | https://www.reddit.com/r/LocalLLaMA/comments/192ccjq/deepseek_coder_33b_vs_dolphin_25_mixtral_8x7b_for/ | oldschooldaw | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 192ccjq | false | null | t3_192ccjq | /r/LocalLLaMA/comments/192ccjq/deepseek_coder_33b_vs_dolphin_25_mixtral_8x7b_for/ | false | false | self | 6 | null |
OCR models in HF? | 1 | Hi All, I'm currently using Azure Vision for getting text with this particularly difficult images: https://imgur.com/a/4ajaMFL
The problem is that I basically need to do for loads of them. I went straight to HF to look for models, but it seems to me there isn't much about this in HF or I'm unable to find it.
Is someone here able to give me directions? Of course traditional software like Tesseract is out of question because it's made to scan documents, not this kind of artsy-images.
I know I'ts a difficult task, but I just need it to recognize the text so I can pass it to a LLM and make something out of it. | 2024-01-09T11:08:19 | https://www.reddit.com/r/LocalLLaMA/comments/192caz8/ocr_models_in_hf/ | iagovar | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 192caz8 | false | null | t3_192caz8 | /r/LocalLLaMA/comments/192caz8/ocr_models_in_hf/ | false | false | self | 1 | {'enabled': False, 'images': [{'id': '3zwH_40bNuKc9D9KS_f8PgYyLB0_StO4UB6QuJ_VuEM', 'resolutions': [{'height': 108, 'url': 'https://external-preview.redd.it/feRmrP1ig8C5cvG7liv5rvlLOrp1VJDAJA38zLtt22A.jpg?width=108&crop=smart&auto=webp&s=66db570353891fad233d5d734db3194a5ceccd47', 'width': 108}, {'height': 216, 'url': 'https://external-preview.redd.it/feRmrP1ig8C5cvG7liv5rvlLOrp1VJDAJA38zLtt22A.jpg?width=216&crop=smart&auto=webp&s=c1613bbc49be9dcabe8181b5ef61c4a0a2a954f8', 'width': 216}, {'height': 320, 'url': 'https://external-preview.redd.it/feRmrP1ig8C5cvG7liv5rvlLOrp1VJDAJA38zLtt22A.jpg?width=320&crop=smart&auto=webp&s=38dcfed1d29a25c3ad5d688ef983d5d2aded9071', 'width': 320}], 'source': {'height': 573, 'url': 'https://external-preview.redd.it/feRmrP1ig8C5cvG7liv5rvlLOrp1VJDAJA38zLtt22A.jpg?auto=webp&s=376982cf7333fe1cdf1bb423951abd3bcf160662', 'width': 573}, 'variants': {}}]} |
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.