Qwen3-0.6B-Dakota-Grammar-RL

Model Description

This model is a reinforcement learning (RL) fine-tuned version of Qwen/Qwen3-0.6B, trained specifically for Dakota language grammar and translation tasks using compositional reward functions on non-coding tasks. This represents a test of the RL pipeline's effectiveness for complex, multi-component reward structures in linguistic domains.

Key Features

• Compositional Rewards: Multi-component reward function combining character preservation, affix accuracy, semantic correctness, pattern matching, and length penalties
• Non-Coding Domain: Demonstrates RL effectiveness beyond code generation tasks
• Dakota Language Focus: Trained on 10,576 grammar tasks extracted from the 1890 Dakota-English Dictionary
• Special Character Preservation: Maintains Dakota orthography (ć, š, ŋ, ḣ, ṡ, á, é, í, ó, ú, etc.)

Training Details

Training Data

• Source: 1890 Dakota-English Dictionary grammar section (pages 1-88)
• Tasks: 10,576 training tasks covering:
  • Morphology (affix application, word formation)
  • Translation (Dakota ↔ English)
  • Reverse translation
  • Syntax (sentence structure)
  • Pattern identification
• Difficulty Levels: Easy (1,973), Medium (5,294), Hard (1,172), Advanced (2,137)

Training Procedure

• Framework: Prime Intellect RL (prime-rl)
• Base Model: Qwen/Qwen3-0.6B
• Training Steps: 1,000 steps
• Batch Size: 256
• Sequence Length: 1,536 tokens
• Rollouts per Example: 8
• Learning Rate: 1e-6
• Checkpoint Interval: Every 100 steps (kept 3 most recent)
• GPUs:
  • Trainer: GPUs 4,5,6,7
  • Inference: GPUs 0,1,2,3

Reward Function Composition

The model was trained using a compositional reward function with the following components:

1. Exact Match Reward (40% weight): Binary reward for exact normalized match
2. Character Overlap Reward (20% weight): F1 score for Dakota special character preservation
3. Pattern Reward (15% weight): Verification pattern matching and hint coverage
4. Affix Reward (10% weight): Accuracy of required morphological affixes
5. Length Penalty Reward (15% weight): Penalizes overly verbose responses (linear decay for responses >3x expected length)

This multi-component approach allows the model to learn nuanced linguistic patterns while maintaining grammatical correctness and orthographic accuracy.

Environment

• Environment: harleycooper/dakota1890 (v0.1.17)
• Framework: Verifiers-compatible RL environment
• Parser: DakotaTranslationParser (preserves Dakota orthography)

Evaluation

Training Metrics

• Final Entropy: 0.2126 (mean), 0.00813 (median)
• Inference Probabilities: 0.87743 (mean), 0.99876 (median)
• Throughput: 7,800 tokens/s
• Model FLOPS Utilization (MFU): 2.6%
• Peak Memory: 11.5 GiB

W&B Run

• Project: dakota-rl-grammar
• Run Name: dakota-0.6b-rl
• View: W&B Run

Intended Use

This model is intended for:

• Research on RL for non-coding linguistic tasks
• Testing compositional reward functions in RL pipelines
• Dakota language grammar and translation tasks
• Demonstrating RL effectiveness beyond code generation domains

Limitations

• Small model size (0.6B parameters) limits capacity for complex grammar rules
• Trained on historical dictionary data (1890) which may not reflect modern Dakota usage
• Limited to single-turn and multi-turn chat formats
• Requires Dakota language knowledge for proper evaluation

Ethical Considerations

• Trained on historical linguistic data from indigenous language documentation
• Should be used respectfully and in consultation with Dakota language communities
• Not intended to replace human language experts or native speakers

Citation

@misc{dakota1890-rl-2024,
  title={Qwen3-0.6B-Dakota-Grammar-RL: A Compositional Reward RL Test for Non-Coding Tasks},
  author={Christian H. Cooper},
  year={2024},
  url={https://huggingface.co/harleycooper/Qwen3-0.6B-Dakota-Grammar-RL}
}

Acknowledgments

• Base model: Qwen/Qwen3-0.6B by Alibaba Cloud
• Training framework: Prime Intellect RL
• Source material: 1890 Dakota-English Dictionary by Stephen Return Riggs
• Environment: Dakota1890 RL environment

Model Card Contact

For questions or issues, please contact: Raise an Issue in the Repo