Manoj2142/taxops-traces

TaxOps: data and agent traces

Synthetic data and the evaluation traces behind every reported number from TaxOps (Lexsi AI Safety & Alignment Hackathon, Track 3): a 3B model (Tiny Aya Global) fine-tuned to file a synthetic federal tax return with six tools while avoiding redundant, failed and decoy-reading calls.

Code: https://github.com/ydv-manoj/taxops-agent · Models: Manoj2142/taxops-tinyaya-sft-v6, Manoj2142/taxops-tinyaya-sft-v3

All data is synthetic, generated deterministically from a case id by taxops/env.py. No real people, employers or tax records.

data/

File	Rows	What
train.jsonl	2,100	training case ids (easy / medium / hard, 700 each) and task prompts
sft.jsonl	2,100	SFT conversations: scripted oracle runs + recovery runs (kind, skip_assistant_turns)
eval.jsonl	300	held-out cases, 100 per tier
eval-shift.jsonl	150	shifted cases: unseen names and unseen decoy types
dev.jsonl / dev-shift.jsonl	30 / 50	checks before full eval runs (hard / shifted), disjoint from eval and eval-shift
sft-v6.jsonl	600	v6 conversations (continued from v3): 535 oracle + 65 bad-id / bad-status recoveries; CORRECTED 1099-INT read in a turn of its own

sft.jsonl is the v4 mix (1,559 oracle + 541 recovery). v3's data came from an earlier version of the generator: run python -m taxops.data export at repository commit 196c8aa to regenerate it.

traces/

One folder per run: traces.jsonl (one episode per line: case id, every tool call with arguments and result, the model messages, the verified score, seconds, tokens; the Claude runs have tool calls only), summary.json (taxops.metrics) and auditkit.json (AuditKIT agent_eval on the same traces).

Run	Cases	Success	Wasted calls / case
base-eval-60 (Tiny Aya Global, untrained)	60	0%	27.8
base-shift-22	22	0%	24.7
sft-v1-eval-30 (first SFT, no calculator in the demos)	30	40%	0.03
sft-v2-eval-30 (the 30-case sample that overstated v2)	30	96.7%	0.03
sft-v2-eval-300 / sft-v2-shift-150	300 / 150	82.7% / 32%
sft-v3-eval-300 (HF A10G, bf16)	300	88.7%	0.06
sft-v3-shift-150 (HF A10G, bf16)	150	68.7%	0.47
sft-v4-eval-300 (recovery demos, degraded)	300	86.0%	0.09
sft-v4-shift-150	150	48.0%	1.13
sft-v3-dev-shift-50 / sft-v6-dev-shift-50 (same Kaggle T4 job)	50	68% / 78%	0.42 / 0.14
sft-v3-eval-300-t4 / sft-v3-shift-150-t4 (v3 rerun on Kaggle T4, fp16)	300 / 150	88.7% / 67.3%	0.057 / 0.47
sft-v6-eval-300 (submitted model, Kaggle T4, fp16)	300	93.3%	0.030
sft-v6-shift-150 (submitted model, Kaggle T4, fp16)	150	74.0%	0.27
claude-sonnet-shift-30 (closed-model baseline)	30	96.7%	0.8
claude-sonnet-eval-30-pilot	30	100%	0.07

Notes:

• Hardware: the original v2–v4 runs used an HF A10G in bf16; v6 and the -t4 v3 reruns used Kaggle T4s in fp16, so compare v6 with the -t4 runs. v3 gave the same outcome on both setups in 300/300 eval and 148/150 shift cases.
• base-shift-22: the run was planned for 30 cases and ran out of GPU memory on case 23; the summary covers the 22 finished cases.
• Claude runs: one fresh Claude Code subagent (Sonnet) per case, allowed to use only the TaxOps CLI (by instruction), from a folder without the source or answers. cli-logs/ holds the raw CLI call logs. The eval pilot reused one subagent across several cases (it could carry memory between them), so only the shift run is used for the comparison.

Rescore any run with the repository's scripts:

PYTHONPATH=. python scripts/merge_shards.py --out /tmp/rescored traces/sft-v6-shift-150   # taxops.metrics summary
python scripts/auditkit_report.py traces/sft-v6-shift-150/traces.jsonl                    # AuditKIT agent_eval

Licence

CC-BY-NC-4.0. The cases are synthetic and generated by MIT-licensed code, but the traces contain outputs of Tiny Aya Global (CC-BY-NC-4.0) and fine-tunes of it, so the dataset follows the same non-commercial terms.