Then I fitted a 10th order poly to each spectrum and patched this over the bad sky lines (always 5577 and sometimes 5894+6300). I likewise patched over 3 bad columns near pixels 100, 140 and 680 in spectrograph 2. Finally I trimmed a few pixels off each end of the spectrum to get rid of rubbish there.
These `corrected' spectra are on the WWW here as fits files. (HINT: this takes about half an hour per spectrograph, about the same as actually doing the z's, so it would be handy if the above all happened inside 2dfdr!!! ) .
Meaning of the Q codes:
Q = 4 is dead-certain z, >= 3 features visible at high significance.
Q = 3 is pretty good, sort of 95% probability that it's right -
e.g. a grotty cross correlation plus a reasonable emission line,
or reasonable Halpha + grotty OII etc.
Q = 2 is sort of 50/50ish - e.g. single line which may be a CR / sky
residual, or a feeble 4000A break etc.
Q = 1 is useless, no hope of getting a z from this spectrum.
I suspect that optimistically maybe half of the Q=2's may become acceptable
if we had better sky subtraction / CR rejection etc,
so the ultimate success rate may go up to 91-92% or so.
The current success rate is 91% for spectrograph 1 and 87% for spec.2; many of the fields were similar for spec1/2, while others e.g. 204, 205, 218, 235, 239 had twice as many failures on spec.2.
The failures are a mixture of `ok' objects, `no object's, LSB's, uncorrected mergers etc. MMC says some of them may be attributable to incorrect fibre/pivot matches (aargh).
Number of stars (defined as Q=3,4 and z < 0.002) is 103 = 2.6% of targets.
Out of the 3445 Q=3/4 redshifts, there were 125 cases `manually' switched from em to abs or vice versa, and another 80 where the eyeball z was neither of the automatic abs/em z.
The semi-auto mode can churn out maybe 400 z's / hour with practice, (the 4's and 1's take only a couple of sec, the 2/3's take maybe 10-20 sec to ponder). so if we want to it is feasible to do the whole 250,000 if we spread the load between 3 or 4 people at 150 hrs each. Or we can do e.g. 20,000 like this and use them as a training set for fully-automatic stuff.
Still to do is redshift error estimates: probably I'll wait until we have a reasonable number of duplicate measurements before looking into this.
eyeball quality means 1=junk, 2=poor, 3=acceptable, 4=good
eyeball override means 0 = auto accepted, 1 = chose the alternate of em/abs, 2 = eyeball z is neither of them.