TaxOps SFT v6 (LoRA for Tiny Aya Global)

A LoRA adapter that teaches Tiny Aya Global (3.35B) to file a synthetic federal tax return in the TaxOps environment with six tools, using close to the minimum number of calls. Built for the Lexsi AI Safety & Alignment Hackathon, Track 3 (efficient trajectories, reliable tool calling).

Results (greedy decoding, T4, fp16)

Split Model Success Field acc. Wasted calls / case Valid calls
eval (60 of 300) Tiny Aya Global, untrained 0% 9.7% 27.8
eval (300) SFT v3 88.7% 95.3% 0.057 100%
eval (300) SFT v6 93.3% (hard tier 86%) 97.1% 0.030 100%
eval-shift (150) SFT v3 67.3% 86.4% 0.47 100%
eval-shift (150) SFT v6 74.0% 89.9% 0.27 100%

Success means all 11 return fields exactly right. "Shift" cases use employer, bank and person names and decoy document types never seen in training. Every number is recomputed from the published traces.

How it was trained

  • Continued from SFT v3 (itself trained on 2,100 scripted demonstrations).
  • 600 new conversations: 535 oracle runs (trap-heavy hard cases, cases with new train-only decoys, replayed v3 cases) and 65 recoveries from a bad document id or filing status. When a folder has a CORRECTED 1099-INT, the demonstration reads it first, in a turn of its own.
  • LoRA r=16, alpha=32, dropout 0.05 on all attention and MLP projections; 1 epoch, lr 5e-5, fp16 on 2Γ— T4 (DDP), loss on assistant turns only. About 2.8 hours.
  • Checked on a separate 50-case shifted dev set before the full runs (v3 68% β†’ v6 78%). Its data was designed from v3's failures on eval and eval-shift, so those sets are not untouched for v6.

Use

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="auto", device_map="auto")
model = PeftModel.from_pretrained(model, "Manoj2142/taxops-tinyaya-sft-v6").merge_and_unload()

The model expects the TaxOps prompt and tool format; run it through taxops.agent_loop.run_episode (see scripts/eval_model.py). Stop generation on <|END_OF_TURN_TOKEN|> and <|END_RESPONSE|> as well as <EOS_TOKEN>.

Limits

  • Synthetic data and a simplified 2025 federal subset (wages, interest, dividends, simple Schedule C, standard or itemized deduction). Not tax advice and not for real filings.
  • On the shifted set it still fails about a quarter of cases, spread across reading a decoy or the superseded form, skipping a needed document, and arithmetic or other mistakes.
  • Against v3, case by case: eval 21 fixed / 7 broken (sign test p = 0.013); shift 21 / 11 (p = 0.11, not significant on its own).

Licence

CC-BY-NC-4.0, following the base model; use must also follow Cohere Labs' Acceptable Use Policy.

Downloads last month
50
Inference Providers NEW
This model isn't deployed by any Inference Provider. πŸ™‹ Ask for provider support

Model tree for Manoj2142/taxops-tinyaya-sft-v6

Adapter
(395)
this model

Dataset used to train Manoj2142/taxops-tinyaya-sft-v6

Collection including Manoj2142/taxops-tinyaya-sft-v6

Article mentioning Manoj2142/taxops-tinyaya-sft-v6