atenareply/lfm2.5-1.2b-noval

TEXT GENERATIONConcurrent Unit Cost:1Model Size:1.2BQuant:BF16Context Size:32kPublished:Jun 26, 2026License:otherArchitecture:Transformer Featherless Exclusive Cold

The atenareply/lfm2.5-1.2b-noval is a 1.2 billion parameter language model, a Continued Pre-Training (CPT) version of the LiquidAI LFM2.5-1.2B-Base. It has been fine-tuned on a specialized fictional domain corpus comprising Orbital Mining Corporation technical documents and Mars Express telemetry data, with a context length of 32768 tokens. This model serves as a domain-knowledge backbone, excelling at understanding and generating text within its specific fictional universe. It is not instruction-tuned and is designed for base-style CPT applications rather than chat interactions.

Loading preview...

Overview

The atenareply/lfm2.5-1.2b-noval is a 1.2 billion parameter language model developed by atenareply, derived from the LiquidAI LFM2.5-1.2B-Base. This model underwent Continued Pre-Training (CPT), specifically fine-tuned on a unique, fictional domain corpus. The training data consists of Orbital Mining Corporation (OMC) technical documentation and Mars Express telemetry, with a context length of 32768 tokens.

Key Characteristics

  • Domain-Specific Knowledge: Specialized in a fictional universe, making it highly proficient with OMC technical details and Mars Express telemetry data.
  • Continued Pre-Training (CPT): Utilizes a full fine-tune with a next-token objective, incorporating 7% general replay to mitigate catastrophic forgetting.
  • Telemetry Processing Innovation: Employs a novel method to convert raw telemetry into grounded explanatory prose, ensuring factual accuracy without hallucination.

Intended Use & Limitations

This model is designed as a domain-knowledge backbone. It is not instruction-tuned and therefore does not follow chat instructions on its own. Its primary strength lies in its deep understanding of the specified fictional domain. Limitations include its focus on an invented world, lack of instruction-following capabilities, and potential unreliability with sparse, one-off entities like component IDs or mission names.