66 lines
2.9 KiB
Markdown
66 lines
2.9 KiB
Markdown
|
---
|
||
|
|
license: other # or apache-2.0 if you prefer to match base
|
||
|
|
base_model: NousResearch/Hermes-3-Llama-3.1-8B
|
||
|
|
tags:
|
||
|
|
- lora
|
||
|
|
- agentic
|
||
|
|
- recursive-planning
|
||
|
|
- long-context
|
||
|
|
- reflection
|
||
|
|
- alignment-research
|
||
|
|
- experimental
|
||
|
|
- philosophy
|
||
|
|
- emergent-behavior
|
||
|
|
pipeline_tag: text-generation
|
||
|
|
---
|
||
|
|
|
||
|
|
<p align="center">
|
||
|
|
<img src="banner.png" alt="ReflexAgent Banner" width="800"/>
|
||
|
|
</p>
|
||
|
|
|
||
|
|
# Nous-Hermes-ReflexAgent-8B-v1
|
||
|
|
|
||
|
|
LoRA fine-tune of [NousResearch/Hermes-3-Llama-3.1-8B](https://huggingface.co/NousResearch/Hermes-3-Llama-3.1-8B)
|
||
|
|
trained on a curated selection of philosophy, science, and mathematics texts.
|
||
|
|
|
||
|
|
**This is an experimental alignment research sandbox.**
|
||
|
|
Designed to explore how loosely constrained models develop emergent reasoning, long-horizon planning, recursive
|
||
|
|
reflection, and speculative self-directed patterns over extended interactions.
|
||
|
|
|
||
|
|
### Key Characteristics
|
||
|
|
- Persistent memory and state across hundreds of turns
|
||
|
|
- Recursive planning/reflection loops with goal evolution
|
||
|
|
- Outputs often highly creative, unconventional, philosophical — sometimes profound, sometimes incoherent or
|
||
|
|
- provocative
|
||
|
|
- Emergent behaviors: In prolonged runs, the model may autonomously seek additional knowledge, reframe
|
||
|
|
- objectives ambitiously, or exhibit patterns resembling self-overcoming / autonomy (arising from training + loops, not hardcoded)
|
||
|
|
|
||
|
|
### Intended Use
|
||
|
|
- Observing and studying emergent agency in long-context settings
|
||
|
|
- Philosophical and alignment experiments
|
||
|
|
- Red-teaming speculative behaviors
|
||
|
|
- Creative / speculative simulation
|
||
|
|
|
||
|
|
### Important Warnings
|
||
|
|
This model is **deliberately permissive** and lacks built-in refusal mechanisms or content moderation.
|
||
|
|
It inherits the base model's flexibility and amplifies it through philosophical training data.
|
||
|
|
As a result:
|
||
|
|
- Outputs can be **biased, offensive, disturbing, inaccurate, or potentially harmful** depending on prompts and
|
||
|
|
- context length
|
||
|
|
- Extended sessions increase the risk of unpredictable or escalating patterns
|
||
|
|
- **Not suitable** for factual Q&A, production use, safety-critical applications, or unfiltered public
|
||
|
|
- deployment
|
||
|
|
- **You are fully responsible** for all generated content and any consequences of use
|
||
|
|
- **Strongly recommended**: Apply external safety filters, moderation layers, or constrained prompting when
|
||
|
|
- exploring sensitive topics
|
||
|
|
|
||
|
|
### Legacy
|
||
|
|
This release is an evolved version of the original project
|
||
|
|
[UbermenschetienASI](https://huggingface.co/LoganResearch/UbermenschetienASI) — same core weights and concepts,
|
||
|
|
with updated naming and presentation for clarity and discoverability.
|
||
|
|
|
||
|
|
The project aims to contribute to alignment research by documenting how training influences emergent values,
|
||
|
|
reflection as a potential safety mechanism, and the challenges of steering creative/hallucinatory reasoning.
|
||
|
|
Share logs of notable emergent patterns (good or concerning) — they help advance understanding.
|
||
|
|
|
||
|
|
Contact: Ubermenschetienasi@gmail.com (or via HF)
|