ziaddddd/granite-3.1-1b-a400m-concision

Overview

This model, granite-3.1-1b-a400m-concision, is an experimental modification of the IBM Granite 3.1 1B-A400M Instruct base model. Its primary purpose is to explore and implement a specific writing-style edit, aiming for increased concision in generated text. The modification involved changing only a single tensor (model.layers.23.self_attn.o_proj.weight) in layer 23, with a strength of 1.0, based on calibration using paired concise/extended prompts.

Key Characteristics & Evaluation

Evaluation was conducted using a partially blind, single-reviewer protocol on a set of 20 test questions. The 'edited' model showed a mean increase of 10.1% in word count compared to the 'original' model across all questions, with a more significant increase (24.9%) for questions requesting detail. This suggests the edit might lead to slightly longer, but potentially more detailed, responses in certain contexts, rather than strictly shorter ones. The model's performance on small screening subsets like GSM8K and ARC_Challenge showed minor changes, with a -6.0 percentage point change on GSM8K and no change on ARC_Challenge, indicating that general capability preservation is not fully established by this limited evaluation.

Artifacts and Limitations

The repository provides the edited Safetensors checkpoint and tokenizer/configuration for Transformers, alongside an F16 GGUF for Ollama. It's important to note that this is an experimental model, and the small screening exercise does not guarantee general capability preservation or production readiness. Observed trade-offs and evaluation limits mean that this model is not intended for production use without further comprehensive testing.