DevQuasar-R1-Uncensored-Llama-8B

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

Eval results

hf (pretrained=DevQuasar/DevQuasar-R1-Uncensored-Llama-8B,parallelize=True,dtype=float16), gen_kwargs: (temperature=0.6,top_p=0.95,do_sample=True), limit: None, num_fewshot: None, batch_size: auto:4 (1,16,64,64)

Tasks	Version	Filter	n-shot	Metric		Value		Stderr
hellaswag	1	none	0	acc	↑	0.6052	±	0.0049
		none	0	acc_norm	↑	0.8021	±	0.0040
leaderboard_bbh	N/A
- leaderboard_bbh_boolean_expressions	1	none	3	acc_norm	↑	0.8360	±	0.0235
- leaderboard_bbh_causal_judgement	1	none	3	acc_norm	↑	0.6043	±	0.0359
- leaderboard_bbh_date_understanding	1	none	3	acc_norm	↑	0.4840	±	0.0317
- leaderboard_bbh_disambiguation_qa	1	none	3	acc_norm	↑	0.6360	±	0.0305
- leaderboard_bbh_formal_fallacies	1	none	3	acc_norm	↑	0.5680	±	0.0314
- leaderboard_bbh_geometric_shapes	1	none	3	acc_norm	↑	0.2760	±	0.0283
- leaderboard_bbh_hyperbaton	1	none	3	acc_norm	↑	0.5440	±	0.0316
- leaderboard_bbh_logical_deduction_five_objects	1	none	3	acc_norm	↑	0.4320	±	0.0314
- leaderboard_bbh_logical_deduction_seven_objects	1	none	3	acc_norm	↑	0.4640	±	0.0316
- leaderboard_bbh_logical_deduction_three_objects	1	none	3	acc_norm	↑	0.6440	±	0.0303
- leaderboard_bbh_movie_recommendation	1	none	3	acc_norm	↑	0.7600	±	0.0271
- leaderboard_bbh_navigate	1	none	3	acc_norm	↑	0.6240	±	0.0307
- leaderboard_bbh_object_counting	1	none	3	acc_norm	↑	0.5440	±	0.0316
- leaderboard_bbh_penguins_in_a_table	1	none	3	acc_norm	↑	0.4658	±	0.0414
- leaderboard_bbh_reasoning_about_colored_objects	1	none	3	acc_norm	↑	0.5640	±	0.0314
- leaderboard_bbh_ruin_names	1	none	3	acc_norm	↑	0.7160	±	0.0286
- leaderboard_bbh_salient_translation_error_detection	1	none	3	acc_norm	↑	0.4920	±	0.0317
- leaderboard_bbh_snarks	1	none	3	acc_norm	↑	0.5899	±	0.0370
- leaderboard_bbh_sports_understanding	1	none	3	acc_norm	↑	0.6880	±	0.0294
- leaderboard_bbh_temporal_sequences	1	none	3	acc_norm	↑	0.2200	±	0.0263
- leaderboard_bbh_tracking_shuffled_objects_five_objects	1	none	3	acc_norm	↑	0.1880	±	0.0248
- leaderboard_bbh_tracking_shuffled_objects_seven_objects	1	none	3	acc_norm	↑	0.1320	±	0.0215
- leaderboard_bbh_tracking_shuffled_objects_three_objects	1	none	3	acc_norm	↑	0.3040	±	0.0292
- leaderboard_bbh_web_of_lies	1	none	3	acc_norm	↑	0.4760	±	0.0316
leaderboard_gpqa	N/A
- leaderboard_gpqa_diamond	1	none	0	acc_norm	↑	0.3232	±	0.0333
- leaderboard_gpqa_extended	1	none	0	acc_norm	↑	0.3498	±	0.0204
- leaderboard_gpqa_main	1	none	0	acc_norm	↑	0.3527	±	0.0226
leaderboard_ifeval	3	none	0	inst_level_loose_acc	↑	0.4628	±	N/A
		none	0	inst_level_strict_acc	↑	0.4365	±	N/A
		none	0	prompt_level_loose_acc	↑	0.3216	±	0.0201
		none	0	prompt_level_strict_acc	↑	0.2902	±	0.0195
leaderboard_math_hard	N/A
- leaderboard_math_algebra_hard	2	none	4	exact_match	↑	0.5798	±	0.0282
- leaderboard_math_counting_and_prob_hard	2	none	4	exact_match	↑	0.2276	±	0.0380
- leaderboard_math_geometry_hard	2	none	4	exact_match	↑	0.1970	±	0.0347
- leaderboard_math_intermediate_algebra_hard	2	none	4	exact_match	↑	0.1036	±	0.0182
- leaderboard_math_num_theory_hard	2	none	4	exact_match	↑	0.3377	±	0.0382
- leaderboard_math_prealgebra_hard	2	none	4	exact_match	↑	0.4715	±	0.0360
- leaderboard_math_precalculus_hard	2	none	4	exact_match	↑	0.1111	±	0.0271
leaderboard_mmlu_pro	0.1	none	5	acc	↑	0.3608	±	0.0044
leaderboard_musr	N/A
- leaderboard_musr_murder_mysteries	1	none	0	acc_norm	↑	0.5920	±	0.0311
- leaderboard_musr_object_placements	1	none	0	acc_norm	↑	0.3867	±	0.0305
- leaderboard_musr_team_allocation	1	none	0	acc_norm	↑	0.3560	±	0.0303

Compare to base DeepSeek-R1-Distill-Llama-8B

Model shows improvements in most if these tests:

Link to eval results

DevQuasar-R1-Uncensored-Llama-8B DeepSeek-R1-Distill-Llama-8B

Merge Details

Merge Method

This model was merged using the Linear merge method.

Models Merged

The following models were included in the merge:

• deepseek-ai/DeepSeek-R1-Distill-Llama-8B
• bunnycore/LLama-3.1-8B-HyperNova-abliteration
• Orenguteng/Llama-3.1-8B-Lexi-Uncensored-V2

Configuration

The following YAML configuration was used to produce this model:

models:
  - model: Orenguteng/Llama-3.1-8B-Lexi-Uncensored-V2
    parameters:
      weight: 0.3
  - model: bunnycore/LLama-3.1-8B-HyperNova-abliteration
    parameters:
      weight: 0.3
  - model: deepseek-ai/DeepSeek-R1-Distill-Llama-8B
    parameters:
      weight: 0.4
merge_method: linear
dtype: float16