Model Request: DeepScaleR-1.5B-Preview

[loretoparisi]

Add support to Berkeley DeepScaleR-1.5B-Preview
Reason: all benchmarks show how this 1.5B model performs better at o1 level in several benchmarks and for its size it is suitable to run on-device.

For what concerns model's weights:

• Total parameters is 1.77B, bits is FP32, file size is 7GB
• Quantization: we quantized at FP16 bits having 3.3GB weight files.

More in details:

• Base Model: Fine-tuned from DeepSeek-R1-Distill-Qwen-1.5B
• Dataset: Approximately 40,000 unique problem-answer pairs compiled from AIME, AMC, Omni-MATH, and Still datasets, including data processing steps
  . Reward Function: Binary system (1 for correct answers, 0 for incorrect or improperly formatted answers)
• Training Steps via RL with GRPO
  Phase 1: 8K Context length (8 samples per prompt)
  Phase 2: 16K Context length (16 samples per prompt)
  Phase 3: 24K Context length