lewisdog/qwen3-1.7b-cogs-ask

Model Overview

lewisdog/qwen3-1.7b-cogs-ask is a 2 billion parameter Qwen3-based model, developed by lewisdog, serving as the student model for the cogs 'ask' pipeline. It is a merged full model (Qwen/Qwen3-1.7B + LoRA) and is specifically engineered to perform two distinct tasks, both outputting compact JSON:

• Decompose: Transforms a single question into 1-4 sub-questions.
• Synthesize: Generates a grounded answer based on retrieved notes, including citations, and reliably abstains from answering out-of-domain questions.

Key Capabilities and Performance

This model demonstrates strong performance in its specialized tasks:

• JSON Output Reliability: Achieves 100% JSON parse rate and 100% non-empty decomposition.
• Citation Handling: Shows 77% citation validity, which can be improved to approximately 90% with a recommended prefix-repair mechanism.
• Abstention Accuracy: Maintains 100% abstention preservation for out-of-domain questions with a low 4% false-abstention rate, making it reliable for declining irrelevant queries.

Training Details

The model was trained using LoRA (r=32/α=64) over 5 epochs with an effective batch size of 16 and a maximum sequence length of 8192. It utilized a cosine learning rate of 1e-4 in bf16 precision. The training data consisted of cogs distill ask pairs, with oversampling for abstention cases.

Usage Notes

For optimal performance, it is recommended to decode at temperature=0 with repetition_penalty=1.1 and to stop generation on <|im_end|>. A specific Python function is provided for repairing citation prefixes to enhance citation validity.