Lambent/qwen2.5-reinstruct-alternate-lumen-14B

qwenreinstruct

This is a merge of pre-trained language models created using mergekit.

Merge Details

Extracted an approximate LoRA of v000000/Qwen2.5-Lumen-14B, rank 128 difference between that and Instruct, and first applied this to Lambent/qwen2.5-14B-alternate-instruct-slerp which had no issues with EQ-Bench.

Then, here, re-applied a density and weight of original Instruct which in previous merges gave me no issues with EQ-Bench.

This one has EQ-Bench of 77.6713 and no "emotions don't match reference error" (if possibly still one not parsed). This is similar to Lumen and original Instruct and slightly exceeds both (within margin of error). My hope is that it has healed Instruct somewhat and regained its intelligence.

Merge Method

This model was merged using the della merge method using Lambent/qwen2.5-lumen-rebased-14B as a base.

Models Merged

The following models were included in the merge:

• Qwen/Qwen2.5-14B-Instruct

Configuration

The following YAML configuration was used to produce this model:

models:
  - model: Qwen/Qwen2.5-14B-Instruct
    parameters:
      weight: 0.3
      density: 0.4
merge_method: della
base_model: Lambent/qwen2.5-lumen-rebased-14B
parameters:
  epsilon: 0.05
  lambda: 1
dtype: bfloat16
tokenizer_source: base

Open LLM Leaderboard Evaluation Results

Detailed results can be found here

Metric	Value
Avg.	33.66
IFEval (0-Shot)	47.94
BBH (3-Shot)	48.99
MATH Lvl 5 (4-Shot)	19.79
GPQA (0-shot)	16.89
MuSR (0-shot)	19.62
MMLU-PRO (5-shot)	48.76