this repo has no description
Bump max_new_tokens to 512 and add raw output debug logging
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
+4
-3
+2
-1
src/app.py
+2
-1
src/app.py
···
50
50
return_tensors="pt",
51
51
).to("cuda")
52
52
with torch.no_grad():
53
-
out = model.generate(**inputs, max_new_tokens=256, do_sample=False)
53
+
out = model.generate(**inputs, max_new_tokens=512, do_sample=False)
54
54
decoded = processor.decode(out[0], skip_special_tokens=False)
55
+
print("[RAW MODEL OUTPUT]\n", decoded, flush=True)
55
56
return extract_assistant(decoded)
56
57
57
58
+1
-1
src/eval.py
+1
-1
src/eval.py
···
70
70
).to("cuda")
71
71
72
72
with torch.no_grad():
73
-
out = model.generate(**inputs, max_new_tokens=256, do_sample=False)
73
+
out = model.generate(**inputs, max_new_tokens=512, do_sample=False)
74
74
75
75
decoded = processor.decode(out[0], skip_special_tokens=False)
76
76
pred = extract_assistant(decoded)
+1
-1
src/probe.py
+1
-1
src/probe.py
···
31
31
return_tensors="pt",
32
32
).to("cuda")
33
33
with torch.no_grad():
34
-
out = model.generate(**inputs, max_new_tokens=256, do_sample=False)
34
+
out = model.generate(**inputs, max_new_tokens=512, do_sample=False)
35
35
decoded_full = processor.decode(out[0], skip_special_tokens=False)
36
36
return extract_assistant(decoded_full)
37
37