hanseungwook/olmo3-recurrent-adapter-sft

OLMo-3 Recurrent Adapter - Answer-Only SFT (rec=1, coda=2, tied)

This is a Recurrent Adapter Model fine-tuned on MetaMathQA for mathematical reasoning, built on top of OLMo-3-1025-7B.

Model Details

• Base Model: allenai/OLMo-3-1025-7B
• Architecture: Recurrent Adapter (1 recurrent layer(s) + 2 coda layers)
• LM Head: tied
• Training Format: Answer-Only
• Mean Recurrence: 32 iterations
• Final Training Loss: 0.1323
• Validation Loss: 3.5968

Training Details

• Dataset: MetaMathQA
• Learning Rate: 1e-4
• Training Steps: 50,000
• Sequence Length: 2048 (Answer-only)
• Gradient Accumulation: 8 steps

Recurrent Adapter Architecture

The model uses a recurrent adapter architecture where:

• The frozen base model extracts initial representations
• A recurrent block (1 layer(s)) processes information over multiple iterations (mean=32)
• A coda block (2 layer(s)) produces the final output
• Only the adapter layers are trained

Usage

import torch
from transformers import AutoTokenizer, AutoModelForCausalLM

model = AutoModelForCausalLM.from_pretrained(
    "hanseungwook/olmo3-recurrent-adapter-sft",
    torch_dtype=torch.bfloat16,
    device_map="auto",
    trust_remote_code=True,
)
tokenizer = AutoTokenizer.from_pretrained("allenai/OLMo-3-1025-7B")

prompt = "What is 25 * 37?"
inputs = tokenizer(prompt, return_tensors="pt").to(model.device)

with torch.no_grad():
    outputs = model.generate(
        **inputs,
        max_new_tokens=512,
        num_recurrence_steps=32,
        temperature=0.7,
    )

print(tokenizer.decode(outputs[0], skip_special_tokens=True))

License

Apache 2.0