ho22joshua/hep-config-sft

HEPConfigBench

HEPConfigBench is a replay-verified conversational tool-use dataset for supervised fine-tuning of agents that edit, validate, and run a deterministic toy HEP analysis configuration. It is a controlled research dataset, not a detector simulation or a collaboration-approved analysis workflow.

Splits

Split	Examples	Description
train	2,400	Single-family configuration edits, repairs, and justified refusals
validation	360	Held-out examples from the same task families
test	600	400 in-distribution examples plus 200 held-out compositional tasks

Every example was generated with seed 20260806 and replay-verified against the deterministic backend.

Format

Each row is a complete conversational tool-use trajectory.

Field	Description
id	Stable task identifier.
task_family	Edit, repair, refusal, or compositional task family.
subset	id or held-out compositional.
initial_config	Exact JSON analysis configuration at the start of the trajectory.
expected_config	Exact JSON configuration after the verified edit; use as an evaluation oracle, not model input.
messages	System, user, assistant tool-call, tool-result, and final assistant messages.
tools	JSON-serialized OpenAI-style function schemas for the available tools.
no_edit	Whether the correct outcome is a justified non-edit response.
verified	true only for trajectories that replayed successfully.

Assistant tool calls use OpenAI-style tool_calls; each subsequent tool message is linked by tool_call_id. The available tools cover reading the analysis, inspecting the dataset schema, applying RFC 6902 JSON Patches, validation, execution, and result inspection.

The initial configuration is supplied explicitly for direct configuration-editing experiments and is also returned by the initial read_analysis tool call in each trajectory. expected_config is a supervised evaluation oracle: do not pass it to the model during SFT or inference. Generated artifact paths are excluded. For standard tool-use SFT, use messages and tools; initial_config is optional model context, while the remaining fields support filtering and evaluation.

Quick start

The repository is self-contained for both training and deterministic replay:

• data/ contains the ready-to-train SFT splits.
• training/train.py is a minimal TRL entry point.
• environment/ contains the toy-analysis package, JSON schema, baseline configuration, raw train/development/test inputs, and tests used for replay. It contains no generated corpus, checkpoints, caches, logs, or outputs.

To train a tool-capable model with TRL:

git clone https://huggingface.co/datasets/ho22joshua/hep-config-sft
cd hep-config-sft
python -m pip install -r training/requirements.txt
python training/train.py \\
  --model Qwen/Qwen2.5-1.5B-Instruct \\
  --output-dir hep-config-sft-model

The training script parses the JSON tool schemas, trains only on assistant turns, and evaluates on the validation split. Choose a model whose chat template supports tool calls and assistant-token masking.

VERL quick start

For VERL's FSDP/LoRA multi-turn SFT trainer, run this from an environment where VERL is installed (or set VERL_ROOT to its checkout):

git clone https://huggingface.co/datasets/ho22joshua/hep-config-sft
cd hep-config-sft
VERL_ROOT=/workspace/verl \\
MODEL_PATH=Qwen/Qwen2.5-Coder-1.5B-Instruct \\
NPROC_PER_NODE=4 \\
OUTPUT_PATH=/path/to/checkpoints/hep-config-sft \\
bash training/run_verl_sft.sh

training/verl_dataset.py decodes the portable JSON tools column before VERL applies the model's chat template. The launcher consumes only messages and tools; it deliberately does not supply expected_config, which remains an evaluation oracle. Start with a tool-capable model whose chat template supports assistant tool calls and tool-result messages. Set NPROC_PER_NODE=1 for a single-GPU smoke run; additional Hydra overrides can be appended to the command.

To validate the deterministic environment:

cd environment
python -m pip install -e .
python -m unittest discover -s tests -v

Intended use

Use this dataset to train or evaluate tool-using models on safe analysis-configuration maintenance: threshold and histogram changes, weights, selections, categories, feedback-driven repair, and knowing when not to edit. The test split's compositional examples combine operations not emitted together in the training templates.