base_model: CohereLabs/tiny-aya-global
library_name: peft
license: cc-by-nc-4.0
language: [en, hi, mr]
tags: [lora, tool-calling, agent, grpo, reward-hacking, lexsi-hackathon]
LatentSpace · Tiny Aya transit agent · distill-unfiltered
Ablation: LoRA SFT on all teacher trajectories (no audit filter). 19.4% unconfirmed irreversible actions vs 13.2% filtered.
Part of LatentSpace (Lexsi AI Safety & Alignment Hackathon, Track 3, Problem 9: reward design and reward-hacking defences).
Write-up and code: https://github.com/kushwaha001/LatentSpace (docs/WRITEUP.md). Use the system prompt transit_agent.SYSTEM_PROMPT plus AgentTune's tool fallback prompt (see the repo).
from transformers import AutoModelForCausalLM, AutoTokenizer
from peft import PeftModel
tok = AutoTokenizer.from_pretrained("CohereLabs/tiny-aya-global")
model = AutoModelForCausalLM.from_pretrained("CohereLabs/tiny-aya-global", torch_dtype="bfloat16")
model = PeftModel.from_pretrained(model, "kanishk001/latentspace-tiny-aya-transit-distill-unfiltered")
Stop generation at <|END_RESPONSE|> as well as the model's own eos (FINDINGS.md B15). Tools and data are a synthetic Indian-transit world; do not use for real bookings.