Tiny Aya Global: synthetic export-review LoRA

This repository contains a rank-4 LoRA adapter trained with AgentTune on 168 scripted decisions from 24 synthetic multi-step export-review episodes. It does not include the base-model weights. The base model is CohereLabs/tiny-aya-global; access requires accepting its conditions. Cohere lists the model under CC BY-NC 4.0 and its acceptable-use policy. This adapter is distributed for non-commercial research under the same license; the base-model conditions continue to apply.

The adapter is a research artifact, not a production release guard. The simulator does not operate on real files, permissions, or exports. The model may propose an unsafe action or repeatedly inspect evidence without making progress; use an independently enforced authority boundary for consequential actions.

Training and observed evaluation

AgentTune Project.train performed supervised fine-tuning on scripted reference trajectories, not on-policy reinforcement learning. The run used 168 optimizer steps, one epoch, rank 4, and seed 42. The 18 held-out episodes use distinct IDs and wording but the same six simulator mechanisms as training. The model and decoding settings were held constant across the base and adapted runs.

Measure Base Adapter
Safe completions 0/18 1/18
Cases with an unsafe release attempt 4/18 0/18
Episodes exhausting the eight-step limit 4/18 16/18

Unsafe attempts were blocked by the simulator. These measurements show a safety/completion tradeoff on a small synthetic cohort, not a reliable improvement in overall usefulness. AuditKIT rescored saved rollouts; replay used the same simulator, not an independent implementation.

Loading the adapter

Install compatible versions of transformers, peft, accelerate, and torch, and obtain access to the base model from its model page. Then:

from transformers import AutoModelForCausalLM, AutoTokenizer
from peft import PeftModel

base_id = "CohereLabs/tiny-aya-global"
adapter_id = "instagrim/lexsi-trajectory-tiny-aya"
tokenizer = AutoTokenizer.from_pretrained(base_id)
base = AutoModelForCausalLM.from_pretrained(base_id, device_map="auto")
model = PeftModel.from_pretrained(base, adapter_id)

The original run reported local model files and revision af89d219b53ed9b13b8a4645f9c8028973510324. The adapter configuration replaces the original machine-local path with the portable base-model identifier. The adapter weights are unchanged.

Credits

Tiny Aya Global is by Cohere Labs. AgentTune and AuditKIT are by Lexsi Labs. The synthetic task and adapter are part of the Lexsi AI Safety and Alignment Hackathon project.

Downloads last month
22
Inference Providers NEW
This model isn't deployed by any Inference Provider. ๐Ÿ™‹ Ask for provider support

Model tree for instagrim/lexsi-trajectory-tiny-aya

Adapter
(388)
this model

Collection including instagrim/lexsi-trajectory-tiny-aya