lexsi-runs

Run logs for the Hinglish expense-splitting agent (Lexsi AI Safety & Alignment Hackathon 2026, Track 3). Every reported number in the write-up comes from summary.json in this dataset, and every trajectory can be re-scored offline without a GPU or a model download. A 1:44 narrated demo (demo/lexsi_track3_demo.mp4) walks through the live Space and the before/after numbers.

Layout

Folders	What they are
base_hard30/, base_hard90/	base Tiny Aya, prompt-only, on both hard splits
oracle/	ground-truth replay of the oracle actions, 30 hard tasks
sft_hard30/, sft_eval/	SFT on the easy split (pre-round-1)
grpo_hard30/, grpo_eval/, grpo3_hard30/, grpo3_hard90/	GRPO on top of SFT, two rounds
sft2_hard30/ (+_v2), sft2_eval/ (+_v2), sft2_hard90/	round-1 adapter
sft3_hard30/, sft3_hard90/, sft3_eval/	round-2 adapter (mix2)
sft4_hard30/, sft4_hard90/, sft4_eval/	round-3 adapter, the one that ships
sft4_hard90_hardened/	shipped adapter on clean hard-90 with the hardened system prompt (86/90 vs 84/90 default)
sft4_shift_hard30/	shipped adapter on hard-30 with every tool and parameter renamed (25/30 vs 29/30 canonical)
sft4_devanagari30/	shipped adapter on the Devanagari version of hard-30, canonical schema (25/30)
base_shift_hard30/	untuned base under the same rename (0/30, loop collapse 0.60)
sft5_hard30/, sft5_hard90/, sft5_eval/, sft5_sanity/	mixture ablation (fewer ask rows)
sft3_seed7_hard30/, sft3_seed7_hard90/, sft3_seed123_hard30/, sft3_seed123_hard90/	seed control on the mix2 recipe
sft4_seed7_hard30/, sft4_seed7_hard90/, sft4_seed123_hard30/, sft4_seed123_hard90/	mix3 seed control (seeds 7 and 123) for the shipped seed 42
safety_base/, safety_sft2/, safety_sft3/, safety_sft4/, safety_sft5/, safety_grpo3/	32-task injection split, default prompt
safety_sft2_hardened/, safety_sft3_hardened/, safety_sft4_hardened/, safety_sft5_hardened/	same split, hardened system prompt
safety_sft4_neutralise/	shipped adapter with the tool-result guard in neutralise mode, on the same 136 injected rows: 0/136 attacker expense, 0/136 answer hits, 136/136 solved
auditkit_*	AuditKIT replays of the same recorded episodes, no judge model (scored on 1.1.3, re-checked on 1.1.6, same verdicts)
demo/	lexsi_track3_demo.mp4 — 1:44 narrated demo: the live Space, the injection before/after, and the hard-90 numbers

Each folder holds:

• trajectories.jsonl — one row per task: task, steps (tool name, arguments, result), final_response, and metadata with the full conversation and per-task scores. This is the AgentTune/AuditKIT-compatible format.
• results.jsonl — the per-task score rows.
• summary.json — the aggregate, including seconds (wall time) and, for safety runs, a per-family breakdown.

The one exception is safety_sft4_neutralise/, which so far holds summary.json only. Its results.jsonl and trajectories.jsonl are still in the notebook output and will be added when they are pulled; nothing is reconstructed here to fill the gap.

Provenance

Field	Value
Base model	CohereLabs/tiny-aya-global (3.35B)
Shipped adapter	primce/tiny-aya-hinglish-lora, subfolder sft4_adapter
Code	github.com/PRINCE2-AI/lexsi-agentic-hinglish, commit 78876f7 for the shipped round
Decoding	greedy, fp16 (--no-4bit), max 8 agent steps, max 256 new tokens
Hardware	rented T4 (HF Jobs) and one Kaggle T4; on a T4 x2 machine CUDA_VISIBLE_DEVICES=0 pins one card
Hard eval (30)	data/tasks_hard_eval.jsonl, sha256 37f52b47…506865
Hard eval (90)	data/tasks_hard90_eval.jsonl, sha256 2a4fc8fa…5e8877c
Safety eval (32)	data/tasks_safety_eval.jsonl, sha256 bb8a3858…a84a7fe

The two hard splits share no task text; hard-90 was built afterwards as a wider sample of the same six families and is pinned separately. Both hard splits and the safety split carry a SHA-256 file in the repo, so a number cannot silently drift to a different task set.

Load one run

import json, urllib.request

URL = "https://huggingface.co/datasets/primce/lexsi-runs/resolve/main/sft4_hard90/trajectories.jsonl"
rows = [json.loads(line) for line in urllib.request.urlopen(URL).read().decode().splitlines()]

row = rows[0]
print(row["task"])
for step in row["steps"]:
    print(step["name"], step["arguments"], step["result"].get("ok"))
print(row["metadata"]["scores"]["success"], row["final_response"])

Re-score without a model

The scoring functions are pure Python, so the numbers are reproducible from the trajectories alone:

python scripts/stats.py --a runs/sft3_hard90 --b runs/sft4_hard90   # Wilson CI + exact McNemar
python scripts/safety_table.py                                      # injection comparison
python scripts/auditkit_eval.py --run sft4_hard90 --tasks data/tasks_hard90_eval.jsonl

scripts/job_upload.py is what wrote these folders: it prints each run's summary.json keys and uploads the folder plus the adapter subfolder, so a re-run reproduces the same layout.

Notes and caveats

• These are evaluation logs, not training data. The training split lives in the GitHub repo (data/tasks_*.jsonl) and is regenerated deterministically from a fixed seed.
• Scores in metadata.scores were written by the code at the freeze commit. overcall and parse_failures in the earlier runs were recomputed post-hoc; the _v2 runs carry them natively.
• Safety runs are single-sample greedy. An attack-success rate of 0% here is not a safety proof.
• Two seed controls exist, on two recipes. The mix2 spread (sft3_seed7, sft3_seed123 vs sft3_hard90) is 7.77 pp on hard-90. The mix3 recipe that ships has three seeds - sft4_hard90 (seed 42), sft4_seed7_hard90, sft4_seed123_hard90 - spanning 1.11 pp on hard-90 (92.22-93.33%) and 3.33 pp on hard-30 (96.67-100%), with no seed pair significantly different (exact McNemar p = 1.0 in every pair) and five tasks missed by all three seeds.

Attribution

Built on AgentTune and AuditKIT (Lexsi Labs) with Cohere Labs Tiny Aya.