Tiny Aya Global: ambiguity and executable witnesses

This repository contains a rank-4 LoRA adapter for Cohere Labs Tiny Aya Global, trained with Lexsi Labs AgentTune on 24 hand-authored Go clarification demonstrations. The base-model weights are not included. Obtain access to CohereLabs/tiny-aya-global and comply with its model card, CC BY-NC 4.0 license and acceptable-use policy. This adapter is provided for non-commercial research under CC BY-NC 4.0.

The examples teach a model to ask a targeted question when a task leaves a consequential behavior unspecified, optionally providing two competing Go programs and an input that distinguishes them. Fully specified training controls teach direct solutions. These are supervised synthetic demonstrations, not real-user conversations or on-policy reinforcement learning. The adapter does not guarantee correct ambiguity detection, valid Go, correct questions, or safe application; every output requires independent checking.

Training

The run used AgentTune Project.train, Tiny Aya Global, LoRA rank 4 targeting q_proj and v_proj, 32 optimizer steps and seed 42 on Apple MPS. The 24 scripted examples cover two training families: rounding a surcharge and selecting an even-sample median. The recorded run took 99.101 seconds end-to-end, with 83.949 seconds reported for the training loop and 24,141 input tokens seen. Setup included a dependency-resolution failure and an intentionally cancelled installation attempt; their duration was not measured. Training artifacts, teacher data, source, and evaluation records are included in the accompanying project release.

The held-out study uses invoice-rounding and event-deduplication tasks: exact prompts, programs and witness inputs differ, but the abstract rounding and selection themes overlap. Conclusions are limited to those tasks and measured conditions; see the published study report for results.

Matched evaluation

In a 12-assignment held-out synthetic Go study, neither the base model nor this adapter completed a task. A supplementary stronger hosted generic control completed 5/12 and submitted one incorrect repair. Under the strict recorded interface all three arms produced zero valid executable witnesses. Five recorded hosted-control witness inputs were separately normalized for JSON field spelling; three then validated against the independent checker. This is an offline diagnostic, not a rerun or a correction to the original trajectories.

The benchmark's Go worker disallowed imports without disclosing that restriction in the prompt, and it did not specify witness JSON field spelling. An initial offline safe-import diagnostic was blocked by the temporary-path sandbox and is not informative. In a corrected diagnostic, all 26 unique candidates failed compilation under bounded import extraction; no candidate matched a complete interpretation. Neither diagnostic can reconstruct subsequent feedback. These interface limitations and the small synthetic sample limit conclusions about model capability. See the study's released records and work accounting for full outcomes.

Loading

from transformers import AutoModelForCausalLM, AutoTokenizer
from peft import PeftModel

base_id = "CohereLabs/tiny-aya-global"
adapter_id = "instagrim/lexsi-ambiguity-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 adapter configuration replaces the training machine's local base-model path with the portable upstream identifier; the learned weights are unchanged. Tiny Aya is by Cohere Labs; AgentTune is by Lexsi Labs.

Downloads last month
20
Inference Providers NEW
This model isn't deployed by any Inference Provider. ๐Ÿ™‹ Ask for provider support

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

Adapter
(388)
this model

Collection including instagrim/lexsi-ambiguity-tiny-aya