Update handler.py
Browse files- handler.py +2 -2
handler.py
CHANGED
|
@@ -5,7 +5,7 @@ Compatible with transformers==4.51.3 (matches model's transformers_version in co
|
|
| 5 |
Generation parameters (from Section A4 of technical review β do not change):
|
| 6 |
do_sample = False (greedy decoding β matches SFT + DPO training exactly)
|
| 7 |
max_new_tokens = 400
|
| 8 |
-
repetition_penalty = 1.
|
| 9 |
no_repeat_ngram_size = 0 (PERMANENTLY DISABLED β hard-coded, not overridable via API)
|
| 10 |
temperature / top_p (REMOVED β inactive under greedy decoding)
|
| 11 |
|
|
@@ -65,7 +65,7 @@ class EndpointHandler:
|
|
| 65 |
# We override it here so model.generate() never reads the repo file.
|
| 66 |
self.model.generation_config = GenerationConfig(
|
| 67 |
do_sample=False, # greedy β matches SFT+DPO training
|
| 68 |
-
repetition_penalty=1.
|
| 69 |
no_repeat_ngram_size=0, # permanently disabled
|
| 70 |
eos_token_id=32007, # <|end|> only β turn terminator
|
| 71 |
pad_token_id=32000, # <|endoftext|>
|
|
|
|
| 5 |
Generation parameters (from Section A4 of technical review β do not change):
|
| 6 |
do_sample = False (greedy decoding β matches SFT + DPO training exactly)
|
| 7 |
max_new_tokens = 400
|
| 8 |
+
repetition_penalty = 1.01 (sole repetition control)
|
| 9 |
no_repeat_ngram_size = 0 (PERMANENTLY DISABLED β hard-coded, not overridable via API)
|
| 10 |
temperature / top_p (REMOVED β inactive under greedy decoding)
|
| 11 |
|
|
|
|
| 65 |
# We override it here so model.generate() never reads the repo file.
|
| 66 |
self.model.generation_config = GenerationConfig(
|
| 67 |
do_sample=False, # greedy β matches SFT+DPO training
|
| 68 |
+
repetition_penalty=1.01,
|
| 69 |
no_repeat_ngram_size=0, # permanently disabled
|
| 70 |
eos_token_id=32007, # <|end|> only β turn terminator
|
| 71 |
pad_token_id=32000, # <|endoftext|>
|