J-shaped, according to "Changing the Risk Paradigms Can be Good for Our Health: J-Shaped, Linear and Threshold Dose-Response Models". A J-shape means the dose-resonse curve is either deflected downward at some point or, stranger yet, starts to demonstrate a reduced risk of the harm of which the toxin is accused. How could that be? Doesn't every toxic exposure produce some finite amount of risk? Nope. 

Assume that a stem cell has some finite lifespan. A toxic exposure followed by whatever random (stochastic) processes are required to complete the transition of the stem cell from normal to malignant must obviously transpire before the cell expires or the cancerous transformation won't occur. But what if the toxic exposure shortens the lifespan of some of the cells? Then you'd expect fewer cells to survive to face the cancer roulette wheel of chance every time they divide.

Unless for some reason the body started producing more cells susceptible to the toxin's effect.

At high levels of exposure the body goes into overdrive regenerating the damaged and destroyed cells. The result is that while some are killed early more cells with higher odds of developing the necessary compliment of mutations/transformations needed for cancer (thanks to priming by the toxin) survive to meet the DNA replication wheel of misfortune. Thus the overall odds that one cell will acquire all the changes needed to initiate cancer increases when exposures are high.

The resultant dose-response curve would show escalating risk above the level where the toxic exposure produces regenerative hyperplasia and a sharp downward deflection in risk below that level. (Imagine a J rotated 180 degrees through its vertical axis and leaned over to the right). And if the effect of low level exposures is indeed to marshal those defenses organized to detect and eliminate random errors of a certain variety then "Voila!" Hormesis - the view that low level exposures can in some cases be preventative, suddenly makes sense (Imagine the J leaned again to the right but now the response down by the hook describes a better sort of response - less cancer).

The paper makes a compelling argument that there are better (which is to say more supportive of stated public health goals and more accurate) models of the dose-response curve than the hoary linear no-threshold version hatched decades ago when doses couldn't reliably be measured and the double helix had yet to be imagined. And it's open source in the bargain. Give it a read.