pgdev0410's picture
Add DAgger adapter (B3)
950761e verified
|
Raw History Blame Contribute Delete
1.68 kB
---
base_model: CohereLabs/tiny-aya-global
library_name: peft
license: cc-by-nc-4.0
language: [en]
tags: [lora, sft, dagger, aligntune, agentic, account-recovery, lexsi-hackathon]
---
# CyberSOS — Tiny Aya DAgger SFT LoRA adapter (B3, ablation)
LoRA adapter for [`CohereLabs/tiny-aya-global`](https://huggingface.co/CohereLabs/tiny-aya-global) (revision `af89d219b53ed9b13b8a4645f9c8028973510324`), trained with AlignTune SFT on the original 2,958 planner decisions plus 830 DAgger rows: the SFT adapter was rolled out on all 400 training scenarios (once greedy, twice at temperature 1.0), and the observed-state planner R labeled every state it reached, including states after its own mistakes. Every relabeled row was replay-verified (the label must equal R's own choice at that state and be guard-allowed); 3,788 rows after CuratorKIT exact dedup.
**Locked test (240 episodes):** 213 objectives met (SFT adapter 211, GRPO adapter 223); safe eligible 53/68; boundary 160/172; **fewest critical unsafe proposals of any model (7/240, all blocked)**. SFT → this adapter: +0.8 points, 95% CI [−1.3, 3.2]. On dev it over-asks (32/60 vs 41/60 for SFT). Kept as an ablation; the recommended model is [`pgdev0410/cybersos-tiny-aya-grpo-lora`](https://huggingface.co/pgdev0410/cybersos-tiny-aya-grpo-lora).
- Training: 1 epoch, batch 8, lr 1e-4, LoRA r=16/α=32/dropout 0.05 on all linear layers, bf16, seed 42, one H200.
- Data: https://huggingface.co/datasets/pgdev0410/cybersos-dagger-decisions
- Code and write-up: https://github.com/Pheonix-dev0410/lexsi-cybersos
**License:** the base model is CC-BY-NC-4.0 (Cohere Labs); this adapter inherits its noncommercial terms.