Questions tagged [large-language-models]
For questions about large language models (LLMs), i.e. language models that are "large" in size and the data they use.
260 questions
0
votes
0
answers
24
views
Rethinking My Deep-Research Agent Workflow — Should We Move Beyond Static Trees? [closed]
I’m reevaluating a deep-research workflow I built earlier and would love some advice.
My previous design used a static tree workflow (fixed width/depth, node = search → extract → summarize → generate ...
- 1
-1
votes
1
answer
46
views
Missing tokenizer.json file - GGUF Conversion [closed]
When converting mistralai/Mistral-Small-3.2-24B-Instruct-2506 to GGUF (via llama_cpp), I get an error saying the tokenizer.json ...
- 99
0
votes
2
answers
104
views
If LLMs like OpenAI / DeepSeek / Gemini exist, why do we still need ML or NLP libraries, now and in the future?
I’m new to AI and NLP, and I’m trying to understand how different tools fit together.
Large Language Models (LLMs) like OpenAI, DeepSeek, or Gemini can already handle many NLP tasks text ...
3
votes
1
answer
114
views
Why do AI language models overuse em dashes compared to human writers?
I've noticed a consistent pattern in AI-generated text: frequent overuse of em dashes (—), sometimes multiple times in a single paragraph. In contrast, in common human writing—even in the sources AI ...
- 131
0
votes
1
answer
75
views
Finding the right setup for a local LLM for big contexts [closed]
i am trying to figure out what i would need for a setup to do the following task:
i have a korean text about 10-20 pages. i need to translate it, anonymize it, and also swap out some words with ...
- 101
2
votes
2
answers
100
views
Why are LLMs for coding not more focused?
When asking llama3.3:70b about its supported natural and programming languages it lists more than a dozen each. As a user I am usually asking questions in one natural language for one programming ...
- 121
0
votes
1
answer
23
views
Does Azure OpenAI or Amazon Bedrock Store the data sent via API calls?
I have some client data that is filled with PII information. I want to use Azure or AWS LLM models, but I am afraid they will use this data for further training or send it to some third party. Could ...
0
votes
0
answers
18
views
How can I use structured JSON data from PostgreSQL to populate LightRAG’s Neo4j graph without letting the LLM hallucinate relationships?
I’m working on a hybrid RAG (Retrieval-Augmented Generation) system that combines:
Structured data from PostgreSQL
A Neo4j graph database
LightRAG for hybrid (graph + vector) search
I want to use ...
0
votes
0
answers
17
views
Are vector distances still relevant if embeddings are created with the same model but different quantization?
I use bge-m3 model to create embeddings and store them to postgres/pgvector.
I am curious if I can:
use F16 quantization during data creation and storage.
then use Q4_K_M quantization for user search/...
- 1
0
votes
1
answer
54
views
Why not scale down over-parametrised models with a low intrinsic dimension
I was just reading the LoRA paper, which states:
We take inspiration from Li et al. (2018a); Aghajanyan et al. (2020)
which show that the learned over-parametrized models in fact reside on
a low ...
- 103
0
votes
0
answers
74
views
Transformer FLOP accounting during forward pass with ISL / OSL
I'm trying to do some accounting for total inference FLOPs for a single request with input/output sequence length $ISL$ and $OSL$, respectively. In particular, I am trying to account for the extra ...
0
votes
1
answer
67
views
What is the longest text that is represented in one token?
We know that LLMs process tokens and a token is 4 characters in average [Source: OpenAI]. There are also tools like the OpenAI Tokenizer which visualize given text.
In German text, there are tokens ...
- 403
4
votes
1
answer
73
views
How to extend LLMs with structured data like a word breakdown dictionary?
For fun I like playing with words, and have collected a lot of structured data on words, such as:
syllables and syllable counts
pronunciations
etc..
One case I was thinking about at one point was ...
- 127
0
votes
0
answers
17
views
Using PDDL for high-level task planning in a quadruped: is it practical, and how to architect the stack?
I’d like to use PDDL for high-level task planning. The motion layer (locomotion, footstep planning, obstacle avoidance) is handled by a low-level controller and local planners. I’m trying to determine ...
- 131
-1
votes
1
answer
61
views
How are very deep models designed?
How did we discover the architecture of state of the art large language models?
- 99