Tito-7B-slerp

Tito-7B-slerp is a merge of the following models using mergekit:

• gordicaleksa/YugoGPT
• mlabonne/AlphaMonarch-7B

🧩 Configuration

slices:
  - sources:
      - model: gordicaleksa/YugoGPT
        layer_range: [0, 32]
      - model: mlabonne/AlphaMonarch-7B
        layer_range: [0, 32]
merge_method: slerp
base_model: mlabonne/AlphaMonarch-7B
parameters:
  t:
    - filter: self_attn
      value: [0, 0.5, 0.3, 0.7, 1]
    - filter: mlp
      value: [1, 0.5, 0.7, 0.3, 0]
    - value: 0.6
dtype: bfloat16

Results

Evaluations on Serbian LLM eval suite (or rather, performance and knowledge of Serbian):

	ARC-E	ARC-C	Hellaswag	BoolQ	Winogrande	OpenbookQA	PiQA	NQ Open	TriviaQA	Avg.
Zamfir-7B	51.85	32.25	46.03	75.59	62.59	26.00	66.81	16.09	36.11	45.92
Mustra-7B	52.95	33.70	45.89	77.55	64.17	30.60	67.25	15.40	34.84	46.93
Tito-7B	55.43	34.73	48.19	77.37	65.27	30.00	67.30	16.7	35.38	47.82
YugoGPT	57.79	34.73	49.89	69.45	64.56	28.20	72.03	15.82	36.14	47.62

Here, all benchmarks were done 0-shot, on the exception of NQ Open and TriviaQA which were done in 5-shot manner, in order to be comparable to Mistral paper.

If we try to replicate OpenLLM Leaderboard results on available Serbian datasets (running an appropriate amount of shots instead of 0), we get:

	ARC	Hellaswag	Winogrande	TruthfulQA	Avg.
Tito-7B	47.27	-	69.93	57.48	58.23
Perucac-7B	49.74	-	71.98	56.03	59.25
YugoGPT	44.03	-	70.64	48.06	54.24
Llama3-8B	42.24	-	61.25	51.08	51.52
SambaLingo	37.88	-	61.48	47.23	48.86

Note that YugoGPT, Llama3 and SambaLingo are all base models, unlike Tito and Perucac.

Open LLM Leaderboard Evaluation Results

Detailed results can be found here

Metric	Tito	YugoGPT
Avg.	70.13	57.34
AI2 Reasoning Challenge (25-Shot)	68.09	58.10
HellaSwag (10-Shot)	86.38	81.44
MMLU (5-Shot)	64.01	60.68
TruthfulQA (0-shot)	57.01	36.60
Winogrande (5-shot)	81.69	76.56
GSM8k (5-shot)	63.61	30.70