ssc-dsai/gc-llm-apertus-70b-instruct-2509

TEXT GENERATIONPricing:Input $3.5 / Cached $0.175 / Output $8.3Concurrent Unit Cost:4Model Size:70BQuant:FP8Context Size:32kPublished:Apr 29, 2026License:apache-2.0Architecture:Transformer0.0K Open Weights Featherless Exclusive Cold

The ssc-dsai/gc-llm-apertus-70b-instruct-2509 is a 70-billion-parameter bilingual (English/French) instruction model developed by Shared Services Canada — Data Science and Artificial Intelligence (SSC-DSAI). It is a LoRA fine-tune of swiss-ai/Apertus-70B-Instruct-2509, adapted to the Government of Canada domain using synthetic question-answer data from the GC Web Archive. This model is optimized for bilingual question answering and summarization grounded in Canadian public web content, and for drafting assistance in public-sector text.

Loading preview...

Model Overview

ssc-dsai/gc-llm-apertus-70b-instruct-2509 is a 70-billion-parameter bilingual (English/French) instruction-tuned model developed by Shared Services Canada — Data Science and Artificial Intelligence (SSC-DSAI). It is an adaptation of the swiss-ai/Apertus-70B-Instruct-2509 base model, fine-tuned using LoRA on synthetic question-answer data derived from the Government of Canada Web Archive, blended with open instruction data to maintain general capabilities. The model was trained entirely on Canadian public-sector infrastructure, ensuring no data left the department.

Key Capabilities

  • Bilingual (English/French) domain adaptation: Specifically fine-tuned for the Government of Canada context.
  • Question Answering & Summarization: Excels at generating responses and summaries grounded in public GC web content.
  • Drafting Assistance: Supports the creation of public-sector text in both official languages.
  • RAG System Integration: Designed to function as a generator within Retrieval-Augmented Generation (RAG) systems for GC documents.
  • Robust Training: Utilizes a diverse training mixture including GC web data (50%), general instruction following (25%), multilingual coverage (10%), math/reasoning (8%), and code retention (7%).

Important Considerations

While optimized for Canadian public-sector use, general capabilities show a slight regression (e.g., ~2 points on MMLU) compared to the base model. Notably, TruthfulQA scores regressed sharply (from 0.605 to 0.507), indicating a potential for confidently incorrect answers in adversarial contexts. The model's knowledge cutoff is around 2025, and it does not cite sources or express calibrated uncertainty. It is not intended for authoritative statements on government policy or automated decision-making affecting individuals.

Intended Use Cases

  • Bilingual question answering and summarization based on Government of Canada public web content.
  • As a generator in RAG systems over GC documents.
  • Drafting assistance for public-sector text in English or French.
  • Research on domain adaptation of open-weight models for Canadian public-sector use.

Finetunes
2 models