SLM-RL-Agent-Data

Companion datasets for the paper Efficiently Enhancing SLM Agents: A Reinforcement Learning Approach to Performance Improvement.

Code	github.com/rezwanh001/slm-rl-agent
Trained models	mr3haque/SLM-RL-Agent
License	Apache-2.0 (this processing); upstream corpora retain their own licenses

This repository bundles the three preprocessed text corpora used to train the entire SLM-RL-Agent framework — a complete three-stage RLHF pipeline (SFT → reward model → PPO) applied to small language models in the 70M–410M parameter range. Each corpus is provided both as a supervised fine-tuning split (sft_train, sft_eval) and a preference-pair split (preference_train, preference_eval) for Bradley–Terry reward-model training.

---

Dataset summary

Dataset	Source	Domain	Disk	SFT train	SFT eval	Pref. train	Pref. eval
tinystories	roneneldan/TinyStories	Short children's stories (ages 3–5 vocabulary)	15 MB	9,499	499	4,749	249
cnn_dailymail	abisee/cnn_dailymail	News articles with human-written summaries	34 MB	9,500	500	4,750	250
wikitext	Salesforce/wikitext (wikitext-103-raw-v1)	Encyclopedic prose from good/featured Wikipedia pages	12 MB	9,500	500	3,996	210

All splits have been deduplicated, prompt-normalized, and truncated to a uniform prompt + response ≤ 512 tokens budget. Preference pairs are synthesised by ranking completions from candidate SLMs with a length/coherence heuristic; the exact pipeline is reproducible via scripts/prepare_all_datasets.py.

---

Quick start

from datasets import load_dataset

# Load TinyStories SFT split
ds = load_dataset("mr3haque/SLM-RL-Agent-Data", name="tinystories", split="sft_train")
print(ds[0])

# Load CNN/DailyMail preference pairs
pref = load_dataset("mr3haque/SLM-RL-Agent-Data", name="cnn_dailymail", split="preference_train")
print(pref[0]["prompt"], "|", pref[0]["chosen"])

Or clone the raw JSON files directly:

huggingface-cli download mr3haque/SLM-RL-Agent-Data \
    --repo-type dataset --local-dir ./slm-rl-data

---

Schema

sft_* splits — list of objects:

{"prompt": "…", "response": "…"}

preference_* splits — list of objects:

{"prompt": "…", "chosen": "…", "rejected": "…"}

Higher-quality continuations are placed in chosen; lower-quality continuations in rejected.

---

How the data is used in the paper

The three corpora are used to produce 15 fully trained RLHF configurations (5 SLM architectures × 3 domains). A reward model is trained on each preference split and then used to PPO-align the corresponding SFT checkpoint.

Stage	Input	Output
SFT	sft_train	LoRA SFT checkpoint
Reward model	preference_train	Bradley–Terry scalar reward model
PPO (RLHF)	SFT checkpoint + reward model + sft_eval prompts	Aligned PPO checkpoint
Evaluation	sft_eval	Perplexity, reward mean/std, ROUGE, BLEU, Distinct-n

All 30 trained checkpoints (15 SFT + 15 PPO) are published in the companion repo mr3haque/SLM-RL-Agent.

---

Citation

@misc{haque2026slmrlagent,
  title  = {Efficiently Enhancing SLM Agents: A Reinforcement Learning Approach to Performance Improvement},
  author = {Haque, Md. Rezwanul},
  year   = {2026},
  howpublished = {\url{https://github.com/rezwanh001/slm-rl-agent}},
  note   = {University of Waterloo, CPAMI Lab}
}

Licensing note

The preprocessing, preference-pair construction, and packaging are released under Apache-2.0. The underlying text in each split is derived from an existing public corpus and remains subject to that corpus's own license — TinyStories (CDLA-Sharing-1.0), CNN/DailyMail (Apache-2.0), and Wikitext-103 (CC BY-SA 3.0). Please consult those upstream licenses before redistribution.