Thinking-Space/UltraData-IF-1.5B

TEXT GENERATIONPricing:Input $0.04 / Cached $0.008 / Output $0.08Concurrent Unit Cost:1Model Size:1.5BQuant:BF16Context Size:32kTool Calling:SupportedPublished:Sep 3, 2026License:otherArchitecture:Transformer Featherless Exclusive Cold

UltraData-IF-1.5B is a 1.78 billion parameter instruction-following model developed by Thinking-Space, initialized from DeepSeek-R1-Distill-Qwen-1.5B. It is trained using the GRPO algorithm on the instruction-following subset of UltraData-SFT-2605, serving as a teacher model for on-policy distillation research. This model is specifically designed for instruction-following tasks and demonstrates improved performance over its base model in various instruction-following benchmarks.

Loading preview...

UltraData-IF-1.5B: An Instruction-Following Teacher Model

UltraData-IF-1.5B is a 1.78 billion parameter language model developed by Thinking-Space, primarily serving as a teacher model for instruction-following experiments in the context of on-policy distillation (OPD) research. It is built upon the DeepSeek-R1-Distill-Qwen-1.5B base model, sharing its Qwen2ForCausalLM architecture, tokenizer, and chat template, which facilitates token-level distillation.

Key Characteristics & Training

  • Base Model: Initialized from DeepSeek-R1-Distill-Qwen-1.5B.
  • Training Method: Utilizes the GRPO (Generalized Reinforcement Policy Optimization) algorithm.
  • Training Data: Fine-tuned on the instruction-following subset of the openbmb/UltraData-SFT-2605 dataset.
  • Primary Domain: Optimized for instruction following tasks.
  • Context Length: Supports a tokenizer model_max_length of 16,384 tokens.

Performance Highlights

As a teacher model, UltraData-IF-1.5B significantly outperforms its student (DeepSeek-R1-Distill-Qwen-1.5B) across various instruction-following benchmarks:

  • Multi-IF (8-language): Achieves 28.58 (final turn) and 40.39 (3-turn mean), compared to the student's 20.84 and 28.56 respectively.
  • Multi-IF (English): Scores 35.54 (final turn) and 45.05 (3-turn mean), against the student's 24.71 and 30.51.
  • IFBench: Demonstrates higher scores in both prompt-level and instruction-level evaluations, with a prompt-level strict score of 17.00 (vs. 10.33 for student).

Use Cases

This model is particularly suitable for researchers and developers interested in:

  • On-policy distillation: Serving as a robust teacher model for distilling instruction-following capabilities into smaller models.
  • Instruction-following tasks: Applications requiring a compact yet effective model for understanding and executing instructions.
  • Research in RLHF and model training: Exploring advanced reinforcement learning techniques like GRPO for language model fine-tuning.

Finetunes
1 models