RL Compositionality - a weizechen Collection

weizechen 's Collections

RL Compositionality

RL Compositionality

updated Oct 17

From f(x) and g(x) to f(g(x)): LLMs Learn New Skills in RL by Composing Old Ones. https://huggingface.co/papers/2509.25123

weizechen/RL-Compositionality-Stage1-RFT-Data

Viewer • Updated Oct 17 • 118k • 53 • 1

Note Stage-1 RFT data from Llama 3.1 8b Instruct. We use this data to fine-tune Llama 3.1 8b Instruct.
weizechen/RL-Compositionality-Stage2-RL-Level1-TrainData

Viewer • Updated Oct 17 • 500k • 39 • 1
weizechen/RL-Compositionality-Stage2-RL-Level2-TrainData

Viewer • Updated Oct 17 • 500k • 41 • 1
weizechen/RL-Compositionality-Stage2-RL-Level8-TestData

Viewer • Updated Oct 17 • 2.05k • 33 • 1

Note The test data containing our string task problem ranges from Level-1 to Level-8. Not intended for training.
weizechen/RL-Compositionality-Stage-1-Model

8B • Updated Oct 17 • 12 • 1

Note The model after Stage 1 training. It is also the base model for Stage 2 for Section 4.1 and Section 4.2.