SoulInPsyAbstract/protocol0-llama-3.1-8b-v5
SoulInPsyAbstract/protocol0-llama-3.1-8b-v5 is an 8 billion parameter Llama 3.1-based instruction-tuned model developed by SoulInPsyAbstract. It is fine-tuned on the protocol0 behavioral compliance dataset to enforce specific constraints like stopping under ambiguity, avoiding fabrication, and concise, single-action responses. This model is designed for specialized behavioral adherence rather than general-purpose instruction following.
Loading preview...
Overview
SoulInPsyAbstract/protocol0-llama-3.1-8b-v5 is a specialized instruction-tuned model based on meta-llama/Llama-3.1-8B-Instruct. It was fine-tuned by SoulInPsyAbstract using LoRA on the protocol0 behavioral compliance dataset (v5, 2,349 examples).
Key Characteristics
This model is distinct from general-purpose instruction-following models as its training focuses on a specific set of behavioral constraints derived from an internal governance protocol (Protocol 0 / CORE LAW). These constraints include:
- Stopping under ambiguity
- Not fabricating unverifiable claims
- Avoiding unsolicited opinions
- Providing single-action responses
- Maintaining conciseness
Training Details
- Base Model:
meta-llama/Llama-3.1-8B-Instruct - Method: LoRA (r=16, alpha=32, dropout=0.05, targeting q/k/v/o_proj modules), with a full merge published.
- Platform: Together AI managed fine-tuning.
- Dataset: 2,349 examples from the
protocol0_sft_v3system prompt.
Current Status
The model has been trained and its weights secured. Behavioral benchmarking against the base model is pending, and users should not treat this as a validated result until a corresponding experiment write-up is published as part of the SIPA OS AI experiment series. The model is distributed under the Llama 3.1 Community License.