Tabish244/YouLearn-V9.1-MiniCPM5-1B-GGUF

YouLearn V9.1 (MiniCPM5 1B) - GGUF

This is the fine-tuned YouLearn V9.1 offline tutor model, quantized to Q4_K_M format for mobile deployment (specifically the YouLearn Android App).

Model Details

• Base Model: Universal OpenBMB MiniCPM5-1B
• Fine-Tuning: Custom YouLearn instructional dataset (V9.1)
• Quantization: Q4_K_M (4-bit medium, optimal for balancing speed, memory, and quality on mobile CPUs/GPUs)
• Primary Use Case: On-device structured educational assistance (Flashcards, Quizzes, Mind Maps, RAG).

Benchmark Summary

The V9.1 model was evaluated in a 24-case production A/B benchmark against the universal base model.

• Strict Formatting & Fallback Compliance: Passes 7/8 hard checks (Base model passes 3/8).
• Inference Speed: ~0.38 seconds avg response time / ~100 tokens/s on target hardware (Base model: ~0.88s avg).
• Structural Integrity: Successfully produces valid JSON for mind maps and enforces source-label discipline.

Known Limitations

• Complex Artifacts Limits: The model may occasionally generate fewer items than requested (e.g., 2 quiz questions instead of 3) due to training biases toward shorter responses.
• Token Cap Issues: Best results are achieved with generation token limits set > 384 to allow complete 8-mark answers and artifact arrays.
• Hallucinations: While it excels at refusing to answer out-of-domain queries ("Not found in the document"), it may hallucinate when provided with exceptionally weak/noisy context rather than falling back.
• Language: Casual Hinglish/Marathi queries outside the study domain may yield sub-optimal structure.

Usage (llama.cpp)

This model can be run using the standard llama.cpp pipeline.

# Basic CLI inference
./llama-cli -m YouLearn-V9.1-MiniCPM5-1B-checkpoint300-Q4_K_M.gguf \
  -p "<|im_start|>system\nYouLearn is a private offline study tutor.\n<|im_end|>\n<|im_start|>user\nExplain photosynthesis.<|im_end|>\n<|im_start|>assistant\n<think>\n\n</think>" \
  -n 512 \
  --top-p 0.9 --temp 0.35 --min-p 0.05

Note: For MiniCPM5, ensure the empty <think>\n\n</think> block is prepended to the assistant's generation to prevent the model from echoing tokens or generating unnecessary thoughts.