VESTA Release

This directory contains the VESTA code, executable environments, training data, and selected checkpoints. Start with the component README for the stage you want to reproduce. The supplied environment classes are executed locally; review them before running generation, training, or evaluation.

Directory map

Path	Contents
code/envs_construction/	API-grounded environment construction, from source ingestion through replay validation and export.
code/generate_workflow/	Replay-validated offline, online, and RL workflow skeleton generators.
code/data_construction/	Offline task naturalization, online user simulation, solver rollout, scoring, and SFT export.
code/sft_train/	LLaMA-Factory SFT configuration and launcher.
code/rl_train/	Local verl/verl-tool runtime, reward implementation, dataset converter, and RL launcher.
code/eval/	Single-checkpoint BFCL, ACEBench, and TauBench v2 launchers and benchmark code.
envs/, data/	Full environment records and released training data.
envs_updated/, data_updated/	Field-minimized copies for downstream use; original files remain in envs/ and data/.
models/SFT/, models/RL/	Selected SFT and RL checkpoints for Qwen3-1.7B, Qwen3-4B, and Qwen3-8B.

Reproduction path

1. Construct executable environments from the external API sources listed in code/envs_construction/md/API_SOURCES.md. Model-assisted construction stages need a separately configured API endpoint.
2. Generate replay-validated skeletons with code/generate_workflow/generate.py: offline (conv) has 16 styles, online (diag) has 16 distinct styles, and RL has 20 styles, including the 16 offline styles.
3. Generate SFT trajectories with code/data_construction/: offline naturalizes fixed user requests before rollout; online uses a user simulator during rollout. The released SFT training file is already in envs_data/vesta_sft_3600_alpha.json.
4. Train SFT with code/sft_train/train.sh and the separately installed LLaMA-Factory revision specified in its README.
5. Convert RL prompt manifests to verl Parquet, then train with code/rl_train/train_sft_rl_{1_7b,4b,8b}.sh. The tool server obtains gold tasks and expected states from a separate lookup; they are never copied into model prompts.
6. Evaluate one checkpoint at a time with the launchers in code/eval/. ACEBench and TauBench require the separately served user simulator (code/eval/user_simulator.sh).

Released data and environments

Use	Environment file	Data file or input
Offline SFT workflows	envs_data/vesta_sft_offline_50envs.jsonl	Offline skeleton generator; released trajectories are in envs_data/.
Online SFT workflows	envs_data/vesta_sft_online_80envs.jsonl	Online skeleton generator; released trajectories are in envs_data/.
SFT training	envs_data/vesta_envs_154.jsonl	envs_data/vesta_sft_3600_alpha.json (Alpaca examples).
RL training	envs_data/vesta_rl_60envs.jsonl	Three 2,400-task training manifests and vesta_rl_15000_gold_lookup.jsonl in envs_data/.

The RL gold lookup contains the 9,002 distinct tasks referenced by the nine released training manifests. Its compact copy preserves the fields required by the converter, tool server, and reward scorer, including each task's reward_components. The larger data/ and envs/ versions are retained for inspection. The 15,000-record RL candidate pool is not itself a training split; the released gold subset does not cover its unused candidate tasks.

The launchers use paths relative to this release directory. RL defaults to the compact files under envs_data/; override VESTA_ENV_ITEMS_PATH and VESTA_MTU_TASK_ITEMS_PATH when using another release layout. The Alpaca SFT file does not carry an env_id; the JSONL trajectories and RL task manifests were checked against their respective environment catalogs.

Quick checks

Run these from this directory; they do not start training or model servers:

python code/generate_workflow/check_catalogs.py
bash code/sft_train/train.sh 4b --dry-run
bash code/rl_train/train.sh 4b --dry-run

Installation, model requirements, full command lines, and benchmark settings are documented in the component READMEs above. The SFT launcher expects LLaMA-Factory; the RL launcher expects its local CUDA/PyTorch stack and a compatible starting SFT checkpoint. Evaluation additionally requires the benchmark-specific dependencies and, for ACEBench/TauBench, the 235B user simulator model.

Current limitation: no released RL eval split

RL evaluation manifests are intentionally not included at present. Although the RL training manifests, gold lookup, and environments are available, the current converter still requires --eval-dir, and the current training launcher and trainer still require a validation Parquet dataset. Therefore the released RL training path does not yet run train-only without an eval input or a subsequent code change. Do not substitute the training set as a validation set and report the resulting numbers as held-out evaluation.

The code and file structure have been checked, but a fresh end-to-end SFT, RL, or benchmark run has not been performed from this release directory.

Publishing layout

A practical public release is a GitHub repository for this README and code/, a Hugging Face Dataset repository for the chosen data_updated/ and envs_updated/ artifacts, and separate Hugging Face Model repositories for the model weights. The full data/ and envs/ copies are useful for internal audit but need not be published alongside the compact copies. Large JSONL files exceed GitHub's ordinary per-file limit; avoid committing them to the code repository. Replace local paths in the component READMEs with the final dataset/model download locations when those repositories exist.

Before publication, review third-party benchmark and vendored-training-code licenses, scan the entire release for credentials and private endpoints, and decide whether the model release should contain inference-ready weights or large optimizer/checkpoint state. This document does not assert that those release checks have been completed.