Add mzansilm probe script (SA-language SLM base validation for the fine-tuning service)

mzansilm_probe.py

@@ -0,0 +1,27 @@
+import json, sys
+print('HF-JOBS-LANE-OK', sys.version.split()[0])
+import torch, transformers
+print('torch', torch.__version__, 'transformers', transformers.__version__)
+from transformers import AutoTokenizer, AutoModelForCausalLM
+MODEL = 'uctnlp/mzansilm-125m'
+tok = AutoTokenizer.from_pretrained(MODEL)
+model = AutoModelForCausalLM.from_pretrained(MODEL)
+model.eval()
+SA = [
+ 'Umuntu wethu umnikazi wosazibeli ngokuqondisa umhlaba wonke umzimba wezokwelapha.',
+ 'Intsapha ye-AI ingenza ulunikezelo lwamanzi ngokuchanekileyo kakhulu.',
+ 'Die plugins moet spoedig ingelewer word.',
+ 'Re a lokwa feela haholo holimo ho ya haholo.',
+]
+tot = 0.0; n = 0
+for s in SA * 3:
+    ids = tok(s, return_tensors='pt').input_ids
+    with torch.no_grad():
+        o = model(ids, labels=ids)
+    tot += float(o.loss); n += 1
+res = {'event':'checkpoint','gate':'mzansilm_probe','status':'recorded',
+       'metrics':{'mean_loss':round(tot/n,4),'samples':n,'vocab':tok.vocab_size,
+                  'params':sum(p.numel() for p in model.parameters())},
+       'model':MODEL,'compute':'hf-jobs-cpu'}
+print('RESULT_JSON=' + json.dumps(res))
+open('probe_result.json','w').write(json.dumps(res) + '\n')