Tiny Aya Global + Borderland A (verified derivations)

A adds checked derivations or provenance reasons to the agent-task targets. GSM8K and Aya rows are identical in A and B.

LoRA (r=32, alpha=64, attention+MLP) on CohereLabs/tiny-aya-global, trained from base with Lexsi AgentTune distillation on 4,003 rows (Borderland 900 + DAgger 121, procedural notes-across-a-wipe tasks 1,272, tool calling 1,060, GSM8K 500, Aya multilingual 150), 2 epochs, lr 1e-4. The A and B adapters use identical rows and hyper-parameters; B's agent-task targets are A's with the reasoning removed; GSM8K and Aya targets are unchanged.

The Borderland chain test uses a teacher-played prefix through room 3, ending immediately before notebook-writing. A then writes its own notebook, is wiped, and opens 58/64 locks (90.6%). This is not an autonomous full-game score. Separate room-4 interventions use empty notes (25%) or the twin world's teacher-written notes (100% follows-twin), not swapped model-written notes. Known weaknesses: over-calls tools on irrelevant requests (BFCL irrelevance 21%), shorter maths reasoning than base.

Held-out results (greedy, single seed, 95% bootstrap CI):

Base A (verified derivations) B (plain traces)
Own notebook -> memory wipe -> open lock after teacher-played prefix (64 worlds, chance 25%) - 90.6% 21.9%
Borderland room success (16x7) 0% 38.4% 6.2%
Note-taking across a wipe, trained task types 2.1% 62.5% 52.1%
BFCL v3 AST (400) 44.5% 54.5% 67.0%
GSM8K (500) 74.8% 57.6% 60.0%
Global-MMLU-Lite, 8 languages (200) 50.5% 54.5% 51.0%

Write-up, data, code and every per-item log: https://huggingface.co/datasets/FatUnc6969/lexi-borderland-data ยท https://github.com/VasuBansal7576/lexi-borderland

from transformers import AutoModelForCausalLM, AutoTokenizer
from peft import PeftModel
base = AutoModelForCausalLM.from_pretrained("CohereLabs/tiny-aya-global", torch_dtype="auto", device_map="auto")
model = PeftModel.from_pretrained(base, "FatUnc6969/tiny-aya-borderland-genA")
tok = AutoTokenizer.from_pretrained("FatUnc6969/tiny-aya-borderland-genA")
Downloads last month
23
Inference Providers NEW
This model isn't deployed by any Inference Provider. ๐Ÿ™‹ Ask for provider support

Model tree for FatUnc6969/tiny-aya-borderland-genA

Adapter
(391)
this model

Collection including FatUnc6969/tiny-aya-borderland-genA