qwen2.5-7b-glyph-sft

Qwen2.5-7B Glyph SFT

Summary

This is a merged supervised fine-tune of Qwen/Qwen2.5-7B-Instruct for the glyph_reasoning research project.

The training setup teaches the model to solve arithmetic and multi-step math problems using a fixed four-part reasoning structure, with special focus on an alchemical "glyph" format:

• 🜞 guideline
• 🜆 plan
• 🜂 step
• 🜃 takeaway / final numeric answer

The goal is research on whether structured symbolic traces can improve reasoning performance and whether those gains transfer beyond the exact training format.

Intended Use

Research use only.

Reasonable uses:

• studying structured reasoning traces
• evaluating glyph-style prompt formats
• comparing format transfer across prompting conditions

Not recommended:

• safety-critical use
• factual QA or broad instruction following beyond the evaluated math setting
• production deployment without additional validation

Training Data

The model was fine-tuned on a synthetic reasoning-trace dataset built from math word problems.

• sources: Math500, GSM8K, and SVAMP style problems
• size: about 3,850 SFT examples
• teacher style: traces produced from structured prompting on the base Qwen model
• output format: chat-style supervised fine-tuning data compatible with the Qwen chat template

Training Procedure

• base model: Qwen/Qwen2.5-7B-Instruct
• total steps: 241
• effective regime: about 1 epoch over the final dataset
• optimizer: AdamW
• schedule: linear decay
• per-device batch size: 1

This uploaded artifact is the merged Hugging Face model, not a LoRA adapter.

Evaluation

The main ablation result for this model comes from a 1000-sample greedy-decoding evaluation on data/unified_dataset.jsonl.

Accuracy

Structure Violation Rate

Interpretation

Main takeaways from the 1000-sample run:

• the SFT model improves over base on glyph
• it also improves strongly on the plain natural baseline
• it gets worse on glyph_shuffled, which suggests sensitivity to the learned structure order
• the emoji result is not clean evidence of format transfer because the model often got the answer while failing the intended emoji structure (1.000 violation rate)

Limitations

• narrow domain: the training and evaluation focus on synthetic math reasoning traces
• mechanism unresolved: the gains may reflect a mix of genuine reasoning transfer and specialization to a specific learned schema
• prompt sensitivity remains: shuffled-symbol performance drops relative to the standard glyph format

Citation

If you use this artifact, cite the base Qwen model and describe it as a fine-tuned research checkpoint from the glyph_reasoning project.