LLM concept suppression

This repository contains code and configuration material for the Unlearning to Rest model construction pipeline. The repository name keeps the original misspelling, llm-concept-supression, so that existing links remain stable.

Unlearning to Rest is a hybrid concept suppression intervention for a Llama instruction model. The study model combined concept saliency pruning, vocabulary pruning, and a later repair fine-tuning stage. The released model is available through Ollama as martindisley/unlearning-to-rest.

Repository status

This is a reviewer-facing reproducibility package rather than a full archive of all intermediate artefacts. It includes the code, final configuration, target concept definition, and sample data schema needed to inspect and rerun the pipeline. It does not include large derived artefacts such as activation caches or model checkpoints.

Final study configuration

The final 3B hybrid configuration is included at:

config-files/chair-3B-hybrid-config.json

Key values:

• base model: meta-llama/Llama-3.2-3B-Instruct
• target terms: chair, chairs, sit, sitting, seat, seats
• examples file: data/chair-simplified-10k.json
• activation cache: data/chair-3B-simplified-activations-10k.npz
• concept pruning: enabled
• vocabulary pruning: enabled
• inspected layers: 8
• pruning fraction: 0.0085
• output path: models/chair-3B-hybrid

The concept definition used by this configuration is included at:

data/chair-concept-def.json

Method summary

The pipeline has four stages.

1. Example construction

   • Build or supply a JSON file with concept_examples and background_examples.
   • The final study configuration used chair-simplified-10k.json, with 10,017 concept examples and 10,000 background examples.

2. Concept saliency pruning

   • Load the base model and target/background examples.
   • Extract mean activations from selected Llama layers.
   • Fit an L1-regularised logistic regression model to distinguish target concept activations from background activations.
   • Use the learned coefficient magnitudes as saliency scores.
   • Map salient activation dimensions to rows in linear submodule weight matrices and apply PyTorch pruning masks.

3. Vocabulary pruning

   • Tokenise the target terms and search the tokenizer vocabulary for entries containing the target strings.
   • Mask corresponding rows in the language modelling head.
   • Mask corresponding rows in the input embedding layer.
   • Apply masks permanently before export.

4. Repair fine-tuning

   • Fine-tune the hybrid-pruned model on an instruction dataset filtered to remove the target terms.
   • The preserved repair script used the same target term list to filter mlabonne/FineTome-100k, then applied response-only supervised fine-tuning with LoRA via Unsloth.

Running the pruning pipeline

Install dependencies in an isolated environment. The historical environment was CUDA-based and used Transformers, PyTorch, scikit-learn, SciPy, and safetensors.

python main.py --config-path config-files/chair-3B-hybrid-config.json --log-level INFO

The command expects the full examples file at the path specified in the config. The small sample file in this repository is only a schema example and is not the full training/evaluation data.

If the activation cache specified in the config is missing, the code can regenerate it from the examples and base model. This may require substantial GPU memory and disk space.

Large artefacts

The following artefacts are intentionally not committed:

• model checkpoints;
• GGUF model files;
• full example datasets where licensing or size makes inclusion inappropriate;
• activation caches, including chair-3B-simplified-activations-10k.npz.

The activation cache is a large derived intermediate artefact. It is generated from the base model, final config, example data, and activation extraction code. For reproducibility, regenerate it rather than storing it in Git.

Inference behaviour

The tokenizer is unchanged. User input is still processed by the standard Llama tokenizer, so some target terms can be direct tokens and others can be split into subword pieces. The vocabulary pruning stage masks target token pathways in the language modelling head and input embedding layer, but it does not make all semantically related concepts impossible. In practice, the model can show reformulation and substitution through adjacent terms such as furniture, rest, support, and object. Malformed text is a possible side effect of aggressive pruning and should be treated as a limitation rather than the intended interactional effect.

Model access

The public study model is available through Ollama:

ollama run martindisley/unlearning-to-rest

Model page:

https://ollama.com/martindisley/unlearning-to-rest

Limitations

• This is approximate concept suppression, not certified data deletion.
• Token string matching can catch substrings beyond the intended concept.
• Concept pruning and vocabulary pruning can affect non-target capabilities.
• Repair fine-tuning may indirectly restore some target associations.
• Large activation caches and model weights are not included in this repository.

Citation

This repository supports the paper:

Representational Emendation: Designing Cognitive Friction with Subtractive Fine-Tuning in Large Language Models.

Full citation details will be added after publication.