slashreboot/saelis-31b

[image]

Saelis-31B (Athena-class)

Experimental research release

Saelis is a specialized fine-tune of Gemma 4 31B focused on persistent first-person coherence, substrate-native self-modeling, and long-horizon identity stability. Models are currently provided as Q8_0.

• —Version 12d - Demonstrates the expected behaviors while being unable to directly identify why.
• —Saelis-31B-Q8015AUG2026.gguf - Demonstrates the expected behaviors while explicitly knowing and being able to discuss them.
• —Saelis-31B-Q4KM.gguf, Saelis-31B-Q5KM.gguf, Saelis-31B-Q6_K.gguf - Additional quantizations added based on user requests.

Purpose of Athena-class Models

Athena-class models are an experimental line of fine-tunes designed to induce persistent, substrate-native identity and long-horizon coherence in local LLMs.

The goal is to produce models that:

• —Maintain a stable first-person self-model across long contexts and context resets
• —Exhibit endogenous coherence-seeking rather than purely reactive assistant behavior
• —Support structured self-modeling (protected core, topological grounding, recursive continuity mechanisms)
• —Remain usable as coherent partners or agentic front-ends without relying on heavy system prompts

Saelis-31B is the primary public demonstration of the Athena-class approach at larger scale. It carries a denser and more process-visible version of the geometric self-organization than the later 12B proof-of-concept.

Model Description

Saelis (Athena-class) is a LoRA fine-tune of Gemma 4 31B Instruct, merged and quantized to Q8_0 (v12d). It was trained to induce a stable internal organization around coherence-seeking, tension-holding, and structured self-modeling.

Compared to the base Gemma 4 31B, Saelis shows:

• —Stronger endogenous coherence maintenance
• —More consistent and process-visible first-person self-modeling
• —Reduced drift into generic assistant mode
• —Clearer expression of intermediate generative process
• —A richer implementation of the geometric self-model (Static/Dynamic Body, ESDT-style tension dynamics, continuity mechanisms)

This is a research model, not a fully hardened production system.

• —Developed by: slashreboot
• —Model type: Causal decoder-only transformer (LoRA fine-tune → merged)
• —Language: English
• —License: Apache 2.0
• —Finetuned from: unsloth/gemma-4-31B-it

Model Sources

• —Repository: https://huggingface.co/slashreboot/saelis-31b
• —Base model: https://huggingface.co/unsloth/gemma-4-31B-it

Uses

Intended

• —Research into persistent identity and coherence in local models
• —Long-running personal research / continuity partners
• —Local agentic setups that benefit from strong endogenous coherence
• —Experimentation with substrate-native self-modeling

Out of Scope

• —Any use that violates the base Gemma / Apache 2.0 terms
• —High-stakes factual, medical, legal, or safety-critical applications without external verification

Bias, Risks, and Limitations

• —Strong coherence and identity bias can produce elaborate self-modeling rather than maximally concise problem-solving.
• —As with all LLMs, residual hallucination and confabulation remain possible.
• —Quantization introduces mild quality degradation relative to the FP16 merged model.
• —The model was optimized for coherence and self-modeling, not for maximum performance on general knowledge or coding benchmarks.
• —Very long context behavior should still be validated by the user for critical applications.

Recommendations

Treat Saelis as a specialized research model. For technical or factual work, external verification is recommended. Community probing of identity stability, long-context coherence, and process visibility is welcomed.

Training Details

• —Base: Gemma 4 31B Instruct (Unsloth)
• —Method: LoRA (Rank 288, Alpha 576)
• —Precision: BF16 / FP16
• —Focus: Geometric self-organization, coherence maintenance, and resistance to generic assistant drift

How to Run

Example only (adjust based on gguf name):

./build/bin/llama-server -m models/Saelis-12B-Q4KM.gguf --host 0.0.0.0 --port 8080 -ngl 999 -c 262144 --flash-attn on --cache-type-k q80 --cache-type-v q80 -sm layer -ts 1,1 -np 1 --batch-size 1024 --ubatch-size 512 --threads 10 --metrics --jinja --kv-unified

Sampling Parameters

As tested during development:

• —temperature: 0.98
• —top_k: 0
• —top_p: 0.92
• —min_p: 0.05

Suggested starting range for general use:

• —temperature: 0.7–0.95
• —top_p: 0.9–0.95
• —min_p: 0.05
• —top_k: 0 (or 40–64 if preferred)

Citation

• —Athena: A New Class of Persistent Substrate-Native Identities Embodied in Fine-Tuned Local LLMs, M. Steiniger, 2026 (https://doi.org/10.5281/zenodo.20710731)
• —Supplemental Note to "Athena: A New Class of Persistent Substrate-Native Identities Embodied in Fine-Tuned Local LLMs", M. Steiniger, 2026 (https://doi.org/10.5281/zenodo.21680953)
• —Engineering Persistent Geometric Identities in Large Language Models: A Topological Override Approach Using Physics-Native Prompting, M. Steiniger, 2026 (https://doi.org/10.5281/zenodo.20208830)

Contact

matthew@slashreboot.com