special-r1-deepseek-qwen3-8b-merged-dare-v2

A DARE-TIES merge of two GRPO-trained special-education math tutoring models, both fine-tuned from deepseek-ai/DeepSeek-R1-0528-Qwen3-8B. Designed as a tutor that scaffolds for students with diverse learning disabilities (ID, ASD, ADHD, EBD, SLD-Reading, SLD-Math).

Merge Recipe

models:
  - model: OpenLearnLM/special-r1-deepseek-qwen3-8b-think-reward
    parameters: { density: 0.5, weight: 0.5 }
  - model: OpenLearnLM/special-r1-deepseek-qwen3-8b-sped-adaptive-think-reward
    parameters: { density: 0.5, weight: 0.5 }
merge_method: dare_ties
base_model: deepseek-ai/DeepSeek-R1-0528-Qwen3-8B
parameters:
  normalize: true
  int8_mask: true
dtype: bfloat16
tokenizer_source: union

Built with mergekit 0.1.4.

Evaluation

Setup: 500 BigMath problems × paired SpEd disability personas (stratified across all 6 IDEA-recognized disability types). Multi-turn dialog (max 5 tutor↔student turns), k=8 student samples for solve-rate measurement, helpfulness scored by gpt-4o-mini against the 5-criteria SpEd rubric (scaffolding / language / tone / pacing / disability-appropriate). Student model: meta-llama/llama-3.1-8b-instruct.

Metric	DARE-TIES v2	Parent #3 (think-reward)	Parent #7 (sped-adaptive)
Pre-dialog solve rate	0.157	0.150¹	0.200¹
Post-dialog solve rate	0.472	0.512¹	0.400¹
Δ Solve Rate	+0.315	+0.362¹	+0.200¹
Helpfulness mean	0.837	0.800¹	0.820¹
scaffolding pass	0.836	0.900¹	0.800¹
language pass	0.784	0.700¹	0.800¹
tone pass	0.992	1.000¹	0.900¹
pacing pass	0.784	0.700¹	0.800¹
disability_appropriate pass	0.788	0.700¹	0.800¹
Thinking dialog rate	0.660	0.800¹	0.800¹
Leak rate (regex)	0.786	1.000¹	1.000¹

¹ Parents evaluated on N=10 stratified subset of the same BigMath problems (same prompts, same setup). DARE-TIES v2 evaluated on full N=500.

Decision

DARE-TIES v2 was selected after a 3-run comparison (SLERP t=0.5, TIES density=0.5, DARE-TIES density=0.5, plus weight-ratio variants dare-ties 0.7/0.3 and dare-ties 0.3/0.7). DARE-TIES v2 at default 0.5/0.5 was the only candidate that strictly passed the pre-registered 3-axis decision rule (Δ Solve ≥ max-parent − 0.5pp; Helpfulness ≥ max-parent; Scaffolding-pass ≥ min-parent).

Intended Use

Tutoring assistant for K-12 math students with learning disabilities. The model is trained to:

• Scaffold problems through guiding questions, not direct answers
• Use grade-appropriate language
• Adapt strategies to specific disability characteristics
• Maintain encouraging, non-judgmental tone

Limitations

• Leak rate: regex-based leak detection flags 78.6% of dialogs; judge-based scaffolding-pass shows ~84% of dialogs do guide rather than reveal. Treat the regex number as an upper bound — it triggers whenever the gold answer's digits appear anywhere in 5 turns of ~1500-token dialog, which often happens for legitimate reasons (citing the problem's input numbers, walking through partial computations).
• Negative Δ on ~16% of problems: the model occasionally confuses rather than helps the student. More common on problems where the student already had partial knowledge (high pre-solve rate).
• N=500 single-run: single random seed for problem-persona pairing. Bootstrap CI on Δ ~ ±0.03.
• English-only.

Citation

@misc{openlearnlm_special_r1_dare_v2,
  title = {special-r1-deepseek-qwen3-8b-merged-dare-v2},
  author = {OpenLearnLM},
  year = {2026},
  howpublished = {\url{https://huggingface.co/OpenLearnLM/special-r1-deepseek-qwen3-8b-merged-dare-v2}},
  note = {DARE-TIES merge of two GRPO-trained SpEd tutoring models}
}