47 lines
1.8 KiB
Markdown
47 lines
1.8 KiB
Markdown
---
|
|
license: llama3.1
|
|
base_model: meta-llama/Llama-3.1-8B-Instruct
|
|
datasets:
|
|
- SoulInPsyAbstract/sipa-os-governance
|
|
tags:
|
|
- fine-tuned
|
|
- lora
|
|
- sipa-os
|
|
- protocol0
|
|
---
|
|
|
|
# protocol0-llama-3.1-8b-v5
|
|
|
|
Fine-tuned `meta-llama/Llama-3.1-8B-Instruct` on the **protocol0** behavioral compliance
|
|
dataset (v5, 2,349 examples), part of the SIPA OS AI experiment series
|
|
(https://github.com/soulinpsyabstract/sipa-os-governance/tree/main/AI_EXPERIMENTS).
|
|
|
|
The dataset trains the model toward a specific set of behavioral constraints — stopping
|
|
under ambiguity, not fabricating unverifiable claims, no unsolicited opinions, single-action
|
|
responses, conciseness — derived from an internal governance protocol (Protocol 0 / CORE LAW),
|
|
not a general-purpose instruction-following upgrade.
|
|
|
|
## Training
|
|
|
|
- Base: `meta-llama/Llama-3.1-8B-Instruct`
|
|
- Method: LoRA (r=16, alpha=32, dropout=0.05, target modules q/k/v/o_proj), full merge published here
|
|
- Platform: Together AI managed fine-tuning
|
|
- Dataset: 2,349 examples, protocol0_sft_v3 system prompt
|
|
(https://github.com/soulinpsyabstract/sipa-os-governance/blob/main/AI_EXPERIMENTS/DATASETS/protocol0_sft_v3_full.jsonl)
|
|
|
|
## Status
|
|
|
|
Trained and weights secured. **Not yet behaviorally benchmarked** against the base model —
|
|
this repo documents both positive and negative/inconclusive results for every experiment in
|
|
this series (see the AI_EXPERIMENTS folder above), and this model's write-up is pending.
|
|
Do not treat this as a validated result until a corresponding EXP write-up is published.
|
|
|
|
## Part of SIPA OS
|
|
|
|
- Full experiment log: https://huggingface.co/datasets/SoulInPsyAbstract/sipa-os-governance
|
|
- Related specialist adapter: https://huggingface.co/SoulInPsyAbstract/specialist-b-refusal-governance
|
|
|
|
## License
|
|
|
|
Distributed under the Llama 3.1 Community License, per the base model's terms.
|