TaxOps SFT v3 (LoRA for Tiny Aya Global)

The first strong TaxOps adapter, and the starting point for the submitted model, SFT v6. Use v6 unless you want to reproduce the v3 → v6 comparison.

It teaches Tiny Aya Global (3.35B) to file a synthetic federal tax return in the TaxOps environment with six tools.

Results (greedy decoding, T4, fp16)

Split	Success	Field acc.	Wasted calls / case	Valid calls
eval (300)	88.7% (easy 100, medium 94, hard 72)	95.3%	0.057	100%
eval-shift (150)	67.3%	86.4%	0.47	100%

The untrained base model solves 0% of the 60 of these eval cases it was run on. An earlier run on an A10G in bf16 gave the same outcome on 300/300 eval and 148/150 shift cases.

How it was trained

2,100 scripted demonstrations (700 each of easy, medium and hard cases); about 20% recover from a bad document id or filing status, with the wrong call masked out of the loss.
The demonstrations do the sums and subtractions with the calculator tool; train-only decoy documents, distinct from the shifted test set's decoys.
LoRA r=16, alpha=32, dropout 0.05 on all attention and MLP projections; 1 epoch, lr 1e-4, bf16 on one A10G, loss on assistant turns only. Data at repository commit 196c8aa.

Use

Same as v6: load the base, apply the adapter with PEFT, and run it through taxops.agent_loop.run_episode (see scripts/eval_model.py).

Limits

Synthetic data and a simplified 2025 federal subset; not tax advice. Its most common failure is the CORRECTED 1099-INT trap (reading the superseded form as well), which v6 halves.