LLM EXPLORER 60,384 MODELS INDEXED

GPT Sw3 356M Instruct by AI-Sweden-Models

By AI-Sweden-Models · 180 downloads

GPT Sw3 356M Instruct is an open-source language model by AI-Sweden-Models. Features: 356m LLM, VRAM: 1.6GB, License: other, Instruction-Based, LLM Explorer Score: 0.11, Arc: 27, HellaSwag: 38, MMLU: 25.5, GSM8K: 1.7.

Base model:ai-sweden-models/gp... Base model:finetune:ai-sweden-...   Conversational   Da Dataset:databricks/databricks-...   Dataset:laion/oig   Dataset:openassistant/oasst1   En   Endpoints compatible   Gpt2   Instruct   Is   No   Pytorch   Region:us   Safetensors   Sv

GPT Sw3 356M Instruct Parameters and Internals

Model Type 
large, decoder-only, transformer, autoregressive
Use Cases 
Areas:
research, evaluation of Large Language Models capabilities
Primary Use Cases:
Validating the model and collecting feedback on Large Language Models.
Limitations:
Bias, Safety, Generation diversity, Hallucination, Overrepresentation of some viewpoints, Discriminatory language, Inaccurate information generation, Repetitive outputs, Content appropriateness, Stereotyping
Considerations:
Awareness of risks and limitations; providing feedback mechanisms to users.
Supported Languages 
da (Danish), sv (Swedish), en (English), no (Norwegian), is (Icelandic)
Training Details 
Data Sources:
databricks/databricks-dolly-15k, laion/OIG, OpenAssistant/oasst1
Data Volume:
320B tokens
Methodology:
Pretrained using causal language modeling with NeMo Megatron GPT implementation. The instruct models were finetuned on instruction data using chat and raw text formats.
Model Architecture:
Large decoder-only pretrained transformer language models.
Safety Evaluation 
Risk Categories:
bias, safety
Ethical Considerations:
Potential for generating biased, incorrect, or harmful content; overrepresentation of some viewpoints.
Responsible Ai Considerations 
Fairness:
Potential bias due to diverse or non-diverse training data.
Transparency:
Increased communication and transparency sought through a modified RAIL license.
Mitigation Strategies:
Encouragement for open communication and feedback collection from indirect users.
Input Output 
Accepted Modalities:
text
Release Notes 
Version:
Second generation
Date:
2022-12-20
LLM NameGPT Sw3 356M Instruct
Repository πŸ€—https://huggingface.co/AI-Sweden-Models/gpt-sw3-356m-instruct 
Base Model(s)  GPT Sw3 356M   AI-Sweden-Models/gpt-sw3-356m
Model Size356m
Required VRAM1.6 GB
Updated2026-07-05
MaintainerAI-Sweden-Models
Model Typegpt2
Instruction-BasedYes
Model Files  1.6 GB   1.6 GB
Supported Languagesda sv en no is
Model ArchitectureGPT2LMHeadModel
Licenseother
Model Max Length2048
Transformers Version4.22.1
Tokenizer ClassGPTSw3Tokenizer
Vocabulary Size64000
Torch Data Typefloat32
Activation Functiongelu