Datasetpublic

atlas-institute/code-trainer-v10-grpo-prompts

code-trainer-v10-grpo-prompts 500 curated prompts for GRPO (Group Relative Policy Optimization) training in the Code-Trainer / RTPI pipeline. Used by both Qwen and Gemma RL stages (Phase 4c) to train tool-call formatting via a rule-based reward function. Schema Column Type Description prompt string The user instruction/question source string Origin: v10_eval, v9_training, or synthetic expected_tool string Primary tool the prompt should invoke… See the full description on the dataset page: https://huggingface.co/datasets/atlas-institute/code-trainer-v10-grpo-prompts.

sourceHugging Faceapache-2.0updated 9d agoView on Hugging Face

0likes40downloads

Dataset Card

code-trainer-v10-grpo-prompts

500 curated prompts for GRPO (Group Relative Policy Optimization) training in the Code-Trainer / RTPI pipeline.

Used by both Qwen and Gemma RL stages (Phase 4c) to train tool-call formatting via a rule-based reward function.

Schema

Column	Type	Description
`prompt`	string	The user instruction/question
`source`	string	Origin: `v10_eval`, `v9_training`, or `synthetic`
`expected_tool`	string	Primary tool the prompt should invoke
`scenario_name`	string	Descriptive scenario label

Sources

Source	Count	Description
V10 eval scenarios	~50	Curated tool-call test cases from `tool_call_eval_entry_v10.py`
V9 training data	~250	User messages preceding tool-call responses, extracted from `code-trainer-v9-mixed`
Synthetic templates	~200	Template-generated prompts covering all 12 Nexus tools

How to use

python

from datasets import load_dataset
ds = load_dataset("cmndcntrlcyber/code-trainer-v10-grpo-prompts", split="train")
print(ds[0]["prompt"])

Reproducibility

bash

python -m src.phase4c_rl.data.build_grpo_prompts --config src/config/pipeline-50.yml