feat: retarget fine-tune pipeline to MiniCPM4.1-8B

- train_modal.py: BASE_MODEL → openbmb/MiniCPM4.1-8B
HUB_REPO → build-small-hackathon/MiniCPM4.1-8B-PaperProf
enable_thinking=False in apply_chat_template
- convert_gguf_modal.py: SRC/GGUF repos updated to 4.1 names
- model/llm.py: DEFAULT_MODEL_ID and GGUF_REPO point to new fine-tune repos

finetune/convert_gguf_modal.py

@@ -1,9 +1,9 @@
 """
 finetune/convert_gguf_modal.py — Convert the fine-tuned model to GGUF for llama.cpp.
 
-Pulls build-small-hackathon/MiniCPM4-8B-PaperProf, converts to GGUF f16 with
+Pulls build-small-hackathon/MiniCPM4.1-8B-PaperProf, converts to GGUF f16 with
 llama.cpp's convert script, quantizes to Q4_K_M, and pushes both files to
-build-small-hackathon/MiniCPM4-8B-PaperProf-GGUF.
+build-small-hackathon/MiniCPM4.1-8B-PaperProf-GGUF.
 
 Run (after the fine-tune has been pushed):
     modal run finetune/convert_gguf_modal.py
@@ -11,8 +11,8 @@ Run (after the fine-tune has been pushed):
 
 import modal
 
-SRC_REPO = "build-small-hackathon/MiniCPM4-8B-PaperProf"
-GGUF_REPO = "build-small-hackathon/MiniCPM4-8B-PaperProf-GGUF"
+SRC_REPO = "build-small-hackathon/MiniCPM4.1-8B-PaperProf"
+GGUF_REPO = "build-small-hackathon/MiniCPM4.1-8B-PaperProf-GGUF"
 
 app = modal.App("paperprof-gguf")
 
@@ -46,7 +46,7 @@ language:
   - en
 ---
 
-# MiniCPM4-8B-PaperProf-GGUF
+# MiniCPM4.1-8B-PaperProf-GGUF
 
 GGUF quantizations of [{SRC_REPO}](https://huggingface.co/{SRC_REPO}),
 the fine-tuned exam-question generator behind
@@ -54,8 +54,8 @@ the fine-tuned exam-question generator behind
 
 | File | Quant | Size | Use |
 |---|---|---|---|
-| `minicpm4-8b-paperprof-Q4_K_M.gguf` | Q4_K_M | ~4.9 GB | recommended, used by the Space |
-| `minicpm4-8b-paperprof-f16.gguf` | F16 | ~16 GB | full precision reference |
+| `minicpm4-1-8b-paperprof-Q4_K_M.gguf` | Q4_K_M | ~4.9 GB | recommended, used by the Space |
+| `minicpm4-1-8b-paperprof-f16.gguf` | F16 | ~16 GB | full precision reference |
 
 ## Usage with llama.cpp
 
@@ -93,8 +93,8 @@ def convert():
     print(f"[pull] downloading {SRC_REPO}…")
     src = snapshot_download(SRC_REPO, token=token, local_dir="/tmp/src")
 
-    f16 = "/tmp/minicpm4-8b-paperprof-f16.gguf"
-    q4 = "/tmp/minicpm4-8b-paperprof-Q4_K_M.gguf"
+    f16 = "/tmp/minicpm4-1-8b-paperprof-f16.gguf"
+    q4 = "/tmp/minicpm4-1-8b-paperprof-Q4_K_M.gguf"
 
     print("[convert] HF → GGUF f16…")
     subprocess.run(

finetune/train_modal.py

@@ -14,9 +14,9 @@ Output:
 
 import modal
 
-APP_NAME = "paperprof-finetune"
-BASE_MODEL = "openbmb/MiniCPM4-8B"
-HUB_REPO = "build-small-hackathon/MiniCPM4-8B-PaperProf"
+APP_NAME = "paperprof-finetune-41"
+BASE_MODEL = "openbmb/MiniCPM4.1-8B"
+HUB_REPO = "build-small-hackathon/MiniCPM4.1-8B-PaperProf"
 N_SAMPLES = 3000
 MAX_LEN = 1024
 
@@ -120,7 +120,7 @@ language:
   - en
 ---
 
-# MiniCPM4-8B-PaperProf
+# MiniCPM4.1-8B-PaperProf
 
 Fine-tuned from [{BASE_MODEL}](https://huggingface.co/{BASE_MODEL}) for
 **exam-question generation** in [PaperProf](https://huggingface.co/spaces/build-small-hackathon/PaperProf),
@@ -326,7 +326,7 @@ def train():
             {"role": "user", "content": pair[0]},
             {"role": "assistant", "content": pair[1]},
         ]
-        return tokenizer.apply_chat_template(messages, tokenize=False)
+        return tokenizer.apply_chat_template(messages, tokenize=False, enable_thinking=False)
 
     texts = [to_text(p) for p in pairs]
 

model/llm.py

@@ -7,8 +7,9 @@ Responsibility:
     about the underlying model loading or tokenisation details.
 
 Model choice:
-    openbmb/MiniCPM4.1-8B — base model, thinking mode disabled for fast
-    structured outputs. Requires transformers >= 4.56.
+    build-small-hackathon/MiniCPM4.1-8B-PaperProf — QLoRA fine-tune of
+    openbmb/MiniCPM4.1-8B on SQuAD/SciQ in PaperProf's production prompt
+    format. Thinking mode disabled. Requires transformers >= 4.56.
 
 Environment variables:
     PAPERPROF_MODEL      Override the default model ID (e.g. "openbmb/MiniCPM3-4B"
@@ -30,7 +31,7 @@ import torch
 from functools import lru_cache
 from transformers import AutoTokenizer, AutoModelForCausalLM, pipeline, BitsAndBytesConfig
 
-DEFAULT_MODEL_ID = "openbmb/MiniCPM4.1-8B"
+DEFAULT_MODEL_ID = "build-small-hackathon/MiniCPM4.1-8B-PaperProf"
 DEFAULT_MAX_NEW_TOKENS = 512
 
 # Pre-load libnvJitLink.so.13 bundled with the nvidia-cu13 wheel so that
@@ -105,7 +106,7 @@ class LLM:
         return output[0]["generated_text"]
 
 
-DEFAULT_GGUF_REPO = "openbmb/MiniCPM4.1-8B-GGUF"
+DEFAULT_GGUF_REPO = "build-small-hackathon/MiniCPM4.1-8B-PaperProf-GGUF"
 
 
 class LlamaCppLLM: