marcodsn/catmind-1.2b

TEXT GENERATIONConcurrent Unit Cost:1Model Size:1.2BQuant:BF16Context Size:32kPublished:Jul 16, 2026License:lfm1.0Architecture:Transformer Featherless Exclusive Cold

marcodsn/catmind-1.2b is a 1.2 billion parameter language model, LoRA-fine-tuned from LiquidAI/LFM2.5-1.2B-Thinking, designed to investigate the role of reasoning trace content. This model replaces actual reasoning steps with short, query-unrelated cat stories while maintaining the original model's final answers during training. It serves as a research artifact to test if the presence of reasoning traces, regardless of their content, is sufficient for performance. The model demonstrates that the content of reasoning traces significantly impacts accuracy, making it primarily suitable for research into LLM reasoning mechanisms.

Loading preview...

marcodsn/catmind-1.2b: A Research Model on Reasoning Traces

marcodsn/catmind-1.2b is a 1.2 billion parameter model, fine-tuned from LiquidAI/LFM2.5-1.2B-Thinking, with a unique approach to its internal <think> block. Instead of generating actual reasoning, this model populates its thinking process with short, irrelevant cat stories, while its final answers were trained to match verified correct solutions from the base model.

Key Findings & Capabilities

This model is a research artifact designed to test whether the content of reasoning traces matters, or if their mere presence is sufficient. Evaluations on 1,000 verifiable-math problems revealed significant insights:

  • Trace Content Matters: Replacing real reasoning with cat stories resulted in a substantial drop in accuracy (from 75.6% to 24.3%) compared to the base model with real reasoning. This indicates that the content within the reasoning trace is crucial for performance.
  • No Hidden Encoding: Prefilling the think block with a random cat story did not further decrease performance, suggesting no hidden reasoning is occurring within the cat stories.
  • Format Consistency: An empty think block further reduced accuracy to 17.2%, implying that any prose, even irrelevant cat stories, provides some benefit in terms of compute or format consistency.

Intended Use & Limitations

  • Research & Entertainment: This model is specifically for research into LLM reasoning mechanisms and for entertainment purposes.
  • Deliberately Worse Performance: It is intentionally worse at math than its base model. It should not be used in scenarios where correct answers are critical.
  • Cat Story Output: The model will always produce a cat story before its answer, regardless of the query.

Finetunes
1 models