SLM-RL-Agents-Data — RLHF Datasets for Small Language Model Agents

Paper: Towards Robust Reinforcement Learning for Small-Scale Language Model Agents

Authors: Md Rezwanul Haque, Md. Milon Islam, Fakhri Karray

Code: github.com/rezwanh001/slm-rl-agents

Models: mr3haque/SLM-RL-Agents

SFT and preference data for three corpora used to train small language model agents via the SLM-RL-Agents RLHF pipeline.

Corpora

Dataset	Domain	SFT Train	SFT Eval	Preference Train	Preference Eval
TinyStories	Narrative fiction	10,000	1,000	10,000	1,000
CNN/DailyMail	News articles	10,000	1,000	10,000	1,000
Wikitext-103	Encyclopaedic	10,000	1,000	10,000	1,000

Repository Layout

SLM-RL-Agents-Data/
└── datasets/
    ├── tinystories/
    │   ├── sft_train.json
    │   ├── sft_eval.json
    │   ├── preference_train.json
    │   └── preference_eval.json
    ├── cnn_dailymail/  (same splits)
    └── wikitext/       (same splits)

Preference Pair Construction

Three complementary degradation strategies in equal proportions:

1. Truncation — chosen = full sentence, rejected = first 50% truncated mid-clause
2. Sentence shuffling — chosen = original order, rejected = random permutation
3. Cross-example mismatch — chosen = genuine reference, rejected = reference from different example

Quick Start

from datasets import load_dataset

ds = load_dataset("mr3haque/SLM-RL-Agents-Data", name="tinystories", split="sft_train")
pref = load_dataset("mr3haque/SLM-RL-Agents-Data", name="cnn_dailymail", split="preference_train")

Citation

@misc{haque2026slmrlagents,
  title   = {Towards Robust Reinforcement Learning for Small-Scale Language Model Agents},
  author  = {Haque, Md Rezwanul and Islam, Md. Milon and Karray, Fakhri},
  year    = {2026},
  howpublished = {\url{https://github.com/rezwanh001/slm-rl-agents}}
}