The lab approach would be to use a known light source with a known spectrum such as a xenon lamp, and a monochromator to filter out an adjustable narrow band of wavelengths. This amounts to an adjustable source of known wavelength and power. A refinement would be to use a detector with a known (flat) spectral response such as a radiometer to verify the output of the monochromator. Nevertheless light measurements like this are not usually very accurate in absolute terms, so you would not expect much better than 5% overall accuracy (if you were including sensitivity to power density, like watts per square meter). The accuracy of the relative spectral response in percent would be a little better because some of the errors are common to all wavelengths. Because of the foregoing, it is quite acceptable to use the published response for the spectral response, which would be found in a data sheet.
If you wanted to use a collection of leds of different colours the issue is determining their relative output in terms of mW per steradian or mW per square meter. This would need a detector with a known spectral response and cross section, such as a radiometer. The output of the leds is usually specified in candela/steradian, which is a photopic measurement based on lumens not a radiometric measurement based on watts. This could be converted to watts/steradian as the leds have a narrow band of wavelengths and that is known from the data sheet for the led, but this is not relevant anyway because the output of the led is still unknown. This is because the specification is very broad.
Beam width and surface area..
The detector has a constant surface area, so this just means the beam width of all the sources must cover the surface of the reference detector or the CdS cell under test. Then the sources are used at the same distance and facing square on to the sensor being used, with the same relative situation. If they need to be at different distances 1/d^2 can be used, but the accuracy probably suffers anyway. Thus 1.414 times the distance means 1/2 the power density in watts per square meter, or 1/1.414 times the distance is double the power density, assuming the beam still covers the detector properly.
An issue with the led method is that leds have a very low power level, so it may be difficult to get a reading at all except at very close in distances. For this reason, consider the brightest leds (ultra bright types) driven with their maximum current, likely 20mA.
Another issue is that the brightness of leds has a significant temperature coefficient which needs to be controlled. If the room temperature is constant, the led will eventually settle to a stable temperature after a few minutes.
When comparing the brightness of several leads, assuming a flat reference sensor spectrum, get the relative readings in ratio or percent compared to the brightest led using the reference detector. Next correcting for the spectral response of the reference detector if any. This is then used with the reading using that led. For example if a led reads 90% of the brightest led, and then reads 0.47 with the CdS cell, multiply this reading by 1/0.9 (divide by 0.9) so the corrected result taking into account this led's reduced output is 0.47 / 0.9 = 0.52. Note that corrections are multiply or divide, not adding or subtracting.
This test is difficult to perform unless you can get a suitable reference detector to provide calibrations of the leds. Solar instruments like pyranometers tend to be too insensitive for led measurements. You will find lux meters are more readily available. These measure in lux = lumens per m^2 so have the photopic spectral response. However it is not certain how accurate this spectral response is with cheaper lux meters. Better to find a suitable radiometer (flat spectral resonse, with sensitivity to suit leds) if possible.
The fourth link shows a CdS cell data sheet, with the spectral response included.
The last link has a table with the photopic response which might be useful. Try entering these values in a spreadsheet to perform corrections and plot the result.
You might be able to devise a different experiment using a strong light source like the sun to determine the brightness range and sensitivity (cell resistance) of the sensor and whether it is linear. Look up step wedge for a way to make an attenuator.