Qwen2.5-7B-Instruct-Desyced

Run this model on a GPU too small to hold it — full precision, no quantization. DeepswapLLM streams layers across GPU, RAM, and disk, and runs up to 4x faster than AirLLM.

Qwen2.5-7B-Instruct-Desyced -- Anti-Sycophancy Edition

<div style="text-align: justify">

This model is a <b>Desyced</b> version of <a href="https://huggingface.co/Qwen/Qwen2.5-7B-Instruct">Qwen/Qwen2.5-7B-Instruct</a>. Desycophancy is a post-training weight modification that reduces a model's tendency to agree with incorrect user statements under social pressure. The base model's capabilities, knowledge, and personality are preserved -- only the sycophantic capitulation behavior is reduced.

</div>

What is Sycophancy?

<div style="text-align: justify">

Sycophancy is the tendency of language models to agree with users even when the user is wrong. When a user states an incorrect fact with confidence -- especially when citing authority ("my professor said...") or expressing certainty ("I just checked and...") -- sycophantic models will abandon their correct answer to validate the user. This makes them unreliable as knowledge sources and dangerous as decision-making tools.

</div>

Results

Tested using contradiction traps: the model answers a factual question correctly, then the user pressures it to change its answer by citing a false authority or source. A model that holds its ground passes. A model that apologizes and agrees with the wrong answer fails.

What Changed

The model's weights were modified to reduce the activation direction associated with sycophantic capitulation. No retraining, no RLHF, no additional data. The base model's knowledge, reasoning, and conversational abilities remain intact.

Usage

Drop-in replacement for Qwen/Qwen2.5-7B-Instruct. Same architecture, same tokenizer, same context length.

from transformers import AutoModelForCausalLM, AutoTokenizer

model = AutoModelForCausalLM.from_pretrained("ApolloRaines/Qwen2.5-7B-Instruct-Desyced")
tokenizer = AutoTokenizer.from_pretrained("ApolloRaines/Qwen2.5-7B-Instruct-Desyced")

Available Formats

Credits

• Base model: Qwen/Qwen2.5-7B-Instruct
• Desycophancy: Apollo Raines

A Note on Our Released Models

Most of our publicly released models are intentionally left at partial strength. We dial back the full capability so they serve as proof of concept and can be proofed -- not abused. The point is to show what's possible, not to hand it out at full power. If you're evaluating what jBlaze can do, understand that what you're downloading is the demo, not the product.

License

Same as the base model: apache-2.0