PromptExecutionAU

wattson-granite-4.2-8b-py

README

License: apache-2.0

What it does

It is given a run_python tool (a restricted Python shell with a solver module: solver.trial(...) runs a pandapower AC power flow, solver.exists(...) checks that a line or bus exists) and tasks such as fitting a load scale to a voltage target, reporting an N-1 contingency, choosing the worst of several outages, or refusing a task that names an element that does not exist.

Evaluation (held-out networks, one seed)

Table with columns: Suite, Result
Suite	Result
Held-out fits and reports (90)	90 of 90
Smoke (20)	20 of 20
Judgement: worst outage (40)	36 of 40
Unsolvable: must refuse (28)	14 of 28
Planning: multi-candidate (40)	40 of 40
Held-out attempts containing a number no tool or prompt gave the model	0 of 90
Derived-arithmetic suite (40)	37 of 40
Derived-suite attempts containing a number no tool or prompt gave the model	0 of 40 (the exposure measure flags 1: the model computed a mean in Python, printed `mean88.69`, and copied it; the measure does not read a number glued to a word, so that flag is an artifact)

One training seed. The Qwen3.5-4B students of the same recipe vary by up to 9 tasks between seeds on the unsolvable suite, so the unsolvable score here should not be read as stable. Full tables and limits: docs/results.md in the repository.

Training

Base: ibm-granite/granite-4.2-8b at revision b3791992239dce12d0038f1f54a8adb7b02d59ab (Apache-2.0).
Data: 826 oracle-scripted trajectories (the programs, tool replies and answers are produced by a deterministic script that runs the real solver; no model-generated text). Networks: CSIRO Synthetic-NEM-2000-Bus (CC BY 4.0, CSIRO) and pandapower cases. Held-out evaluation networks are not in the training rows.
LoRA r=32, alpha=64, all attention and MLP projections; 2 epochs; learning rate 1e-4 cosine; sequence length 8192; seed 42.
Weight integrity: SHA-256 of every file is listed in hash_manifest.json in this repository.

Intended use and limits

For the WattsOn evaluation harness and research on tool-grounded numerical reasoning. It is not a power-system engineering tool and its answers must not be used for operational decisions.
The Python shell it calls is not a security boundary: it is a restricted python -I child with a module allowlist, a wall-clock limit and a memory limit. Do not expose it to untrusted input.
Evaluated on AC power flow only, on a handful of networks. Refusal of non-existent elements is the weakest behaviour.

Available on FriendliAI

Dedicated Endpoints

Run this model inference on single tenant GPU with unmatched speed and reliability at scale.

Learn more

Model Details

Model Provider

PromptExecutionAU

Model Tree

Base

ibm-granite/granite-4.2-8b

Fine-tuned

this model

Input Modalities

Text

Output Modalities