Eklav-4B-Math
Eklav trains a model to pick up a teacher's reasoning mid thought rather than imitate it end to end. The student sees a partial reasoning trace from the teacher, with the answer revealing tail removed, and learns to continue reasoning and produce the answer on its own. The model's own reasoning is conditioned on the teacher's partial trace during training rather than trained to reproduce it word for word. Same base model, same training data as standard full trace CoT distillation, only the training objective changes.
Highlights
- +3.5% average pass@1 across 6 math benchmarks vs. standard full trace CoT SFT, same base model and training data (MMLU excluded, general knowledge check not a math benchmark)
Model details
| Base model | Qwen/Qwen3-4B |
| Task | Math reasoning |
| Training method | Eklav (hint conditioned SFT) |
| Format | Merged bf16 checkpoint |
Results
Pass@1 (%), single evaluation run per benchmark.
| Benchmark | AIME 2025 | AIME 1983-2024 | GPQA-Diamond | GSM8K | MATH-500 | Omni-MATH | MMLU |
|---|---|---|---|---|---|---|---|
| CotGen | 31.9 | 52.6 | 46.0 | 93.8 | 88.7 | 34.7 | 78.7 |
| Eklav | 36.0 | 53.0 | 47.3 | 94.3 | 91.0 | 38.2 | 78.4 |
Usage
from transformers import AutoModelForCausalLM, AutoTokenizer
model_id = "AdarshSingh7647/Eklav-4B-Math"
tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(model_id, torch_dtype="auto", device_map="auto")