LFM3-1.7B-GMCE (GMCoder Efficiency)
This is LiquidAI/LFM3-1.7B fine-tuned for GMCoder Efficiency, which reduces the number of tokens the model spends reasoning before it answers.
Files: GGUF (Q8_0, Q6_K, Q5_K_M, Q4_K_M, BF16) and BF16 safetensors.
The GGUF files need llama.cpp b11960 or newer (lfm3 support). The files were made and tested with b12078. Run llama-server with --jinja. The safetensors use the base model's remote code, so load them with trust_remote_code=True.
License: LFM Open License v1.0, inherited from the base model (see LICENSE).
Measurements
- Exams: three of our own unseen "Form B" exams, with a 60k token cap and context 65536.
- Sampling: T 0.6 / top_p 0.95 / top_k 20, both models at their default chat-template setting (thinking on).
- Both models ran as Q8_0.
- One run per model and one grader. Scores are approximate, so differences of about 5 points are within noise.
- The grader had seen the base answers before grading, so the grading was not fully blind.
| Exam | Base score | GMCE score | Base tokens | GMCE tokens |
|---|---|---|---|---|
| Ultimate (25 math/logic/CS questions) | 17 / 25 | 14.5 / 25 | 43,620 | 18,112 (−58%) |
| Historical (20 engineering questions, approx /100) | ~52 | ~38.5 | 13,970 | 5,405 (−61%) |
| Everyday (30 coding questions, approx /150) | ~88 | ~66.5 | 25,184 | 6,308 (−75%) |
| Total tokens | 82,774 | 29,825 (−64%) |
- Both models finished all three exams inside the cap.
- Wall time for all three exams on an RTX 5070 Ti: base about 6.5 min, GMCE about 2.5 min.
- Most of the token reduction is in the reasoning section. The final answers are also shorter.
Training
- Method: QLoRA (NF4), LoRA r8 / alpha 8 on all linear layers (153 modules, 8.2M trainable params).
- Hyperparameters: LR 5e-5, batch 1, 16 epochs (8,496 steps), completion-only loss, thinking-off render.
- Data: the 531-example GMCoder corpus.
- The adapter was merged into the original BF16 base weights.