OVRLab/granite-3.1-1b-a400m-concision-experiment
OVRLab/granite-3.1-1b-a400m-concision-experiment is an experimental 1 billion parameter language model derived from IBM Granite 3.1 1B-A400M Instruct, developed by OVRLab. This model features a modified attention weight matrix intended to reduce response verbosity, though the experiment showed only a small, inconsistent reduction in length. It serves as a worked example for model editing and reporting weak results, rather than a recommended configuration for general use. The model maintains the original 32768 token context length and exhibits no measured gains or regressions on small GSM8K and ARC-Challenge subsets.
Loading preview...
OVRLab/granite-3.1-1b-a400m-concision-experiment: A Concision Experiment
This model is an experimental result from OVRLab, based on the IBM Granite 3.1 1B-A400M Instruct model. It features a targeted modification to one attention weight matrix (specifically model.layers.12.self_attn.o_proj.weight) with the goal of reducing unnecessary verbosity in responses.
Key Findings & Capabilities
- Experimental Edit: A norm-preserving directional style edit was applied to reduce response length, calibrated on 16 paired concise/extended-answer prompts.
- Limited Impact on Concision: The edit resulted in a small, inconsistent reduction in average response length (2.85% shorter), with 6 shorter, 8 equal-length, and 6 longer responses across 20 questions. In contrast, a simple "concise instruction" prompt reduced length by 25.48% on the original model.
- Capability Preservation: On small 50-question subsets of GSM8K and ARC-Challenge, the edited model showed no change in correct answers compared to the original, indicating preservation of these specific capabilities. However, this does not establish general capability preservation.
- Model Details: The base model is IBM Granite 3.1 1B-A400M Instruct, with approximately 1.3 billion total parameters and 400 million active parameters. The experiment did not involve gradient updates and focused solely on a single weight modification.
Intended Use & Limitations
This release is primarily intended for inspecting and reproducing a narrowly scoped model-editing experiment. It serves as an illustration of how to report weak results and compare them with simple prompting controls. It is not a recommended configuration for production use due to the inconsistent concision improvements and the limited scope of evaluation. The original model's limitations and potential factual errors persist, and no comprehensive safety or multilingual assessments were performed.