Mehdi commited on
Commit
5ad43b3
·
1 Parent(s): 7458c71

feat: upgrade to MiniCPM4.1-8B + official GGUF, disable thinking mode

Browse files

- DEFAULT_MODEL_ID → openbmb/MiniCPM4.1-8B
- DEFAULT_GGUF_REPO → openbmb/MiniCPM4.1-8B-GGUF
- enable_thinking=False in apply_chat_template to suppress chain-of-thought
tokens and keep structured question/evaluation outputs clean
- transformers==4.57.1 satisfies the new >=4.56 requirement, no pin change needed

Files changed (1) hide show
  1. model/llm.py +6 -7
model/llm.py CHANGED
@@ -7,10 +7,8 @@ Responsibility:
7
  about the underlying model loading or tokenisation details.
8
 
9
  Model choice:
10
- build-small-hackathon/MiniCPM4-8B-PaperProf — our QLoRA fine-tune of
11
- openbmb/MiniCPM4-8B, trained on SQuAD passage/question pairs formatted
12
- with the production prompt so it is optimized for exam-question
13
- generation over course material. Requires transformers >= 4.50.0.
14
 
15
  Environment variables:
16
  PAPERPROF_MODEL Override the default model ID (e.g. "openbmb/MiniCPM3-4B"
@@ -32,7 +30,7 @@ import torch
32
  from functools import lru_cache
33
  from transformers import AutoTokenizer, AutoModelForCausalLM, pipeline, BitsAndBytesConfig
34
 
35
- DEFAULT_MODEL_ID = "build-small-hackathon/MiniCPM4-8B-PaperProf"
36
  DEFAULT_MAX_NEW_TOKENS = 512
37
 
38
  # Pre-load libnvJitLink.so.13 bundled with the nvidia-cu13 wheel so that
@@ -95,7 +93,8 @@ class LLM:
95
  """Run *prompt* through the model and return the generated text only."""
96
  messages = [{"role": "user", "content": prompt}]
97
  text = self._tokenizer.apply_chat_template(
98
- messages, tokenize=False, add_generation_prompt=True
 
99
  )
100
  output = self._pipe(
101
  text,
@@ -106,7 +105,7 @@ class LLM:
106
  return output[0]["generated_text"]
107
 
108
 
109
- DEFAULT_GGUF_REPO = "build-small-hackathon/MiniCPM4-8B-PaperProf-GGUF"
110
 
111
 
112
  class LlamaCppLLM:
 
7
  about the underlying model loading or tokenisation details.
8
 
9
  Model choice:
10
+ openbmb/MiniCPM4.1-8B — base model, thinking mode disabled for fast
11
+ structured outputs. Requires transformers >= 4.56.
 
 
12
 
13
  Environment variables:
14
  PAPERPROF_MODEL Override the default model ID (e.g. "openbmb/MiniCPM3-4B"
 
30
  from functools import lru_cache
31
  from transformers import AutoTokenizer, AutoModelForCausalLM, pipeline, BitsAndBytesConfig
32
 
33
+ DEFAULT_MODEL_ID = "openbmb/MiniCPM4.1-8B"
34
  DEFAULT_MAX_NEW_TOKENS = 512
35
 
36
  # Pre-load libnvJitLink.so.13 bundled with the nvidia-cu13 wheel so that
 
93
  """Run *prompt* through the model and return the generated text only."""
94
  messages = [{"role": "user", "content": prompt}]
95
  text = self._tokenizer.apply_chat_template(
96
+ messages, tokenize=False, add_generation_prompt=True,
97
+ enable_thinking=False,
98
  )
99
  output = self._pipe(
100
  text,
 
105
  return output[0]["generated_text"]
106
 
107
 
108
+ DEFAULT_GGUF_REPO = "openbmb/MiniCPM4.1-8B-GGUF"
109
 
110
 
111
  class LlamaCppLLM: