Open-Weight Models

STRIDE with nonnegative local step-diversity credit on eligible reasoning tokens. Training explicitly renders enablethinking=False, recorded as thinkingmode: false in the scientific contract. The pinned tokenizer and chat template remain unchanged. Use the same explicit keyword at inference, particularly for Qwen3-1.7B, whose default template enables thinking. Every checkpoint in this repository requires explicit nonthinking provenance. This public repository retains every published optimizer-update adapter, including update zero (the initial, untrained adapter). Training is planned for 4 epochs on the same 2,048-question training split as the earlier STRIDE runs. The global prompt batch is…

STRIDE with nonnegative local step-diversity credit on eligible reasoning tokens. Training explicitly renders enablethinking=False, recorded as thinkingmode: false in the scientific contract. The pinned tokenizer and chat template remain unchanged. Use the same explicit keyword at inference, particularly for Qwen3-1.7B, whose default template enables thinking. Every checkpoint in this repository requires explicit nonthinking provenance. This public repository retains every published optimizer-update adapter, including update zero (the initial, untrained adapter). Training is planned for 4 epochs on the same 2,048-question training split as the earlier STRIDE runs. The global prompt batch is…

STRIDE ablation using one pooled reasoning embedding per eligible response. Training explicitly renders enablethinking=False, recorded as thinkingmode: false in the scientific contract. The pinned tokenizer and chat template remain unchanged. Use the same explicit keyword at inference, particularly for Qwen3-1.7B, whose default template enables thinking. Every checkpoint in this repository requires explicit nonthinking provenance. This public repository retains every published optimizer-update adapter, including update zero (the initial, untrained adapter). Training is planned for 4 epochs on the same 2,048-question training split as the earlier STRIDE runs. The global prompt batch is 64…

STRIDE ablation distributing each response diversity bonus uniformly over its eligible tokens. Training explicitly renders enablethinking=False, recorded as thinkingmode: false in the scientific contract. The pinned tokenizer and chat template remain unchanged. Use the same explicit keyword at inference, particularly for Qwen3-1.7B, whose default template enables thinking. Every checkpoint in this repository requires explicit nonthinking provenance. This public repository retains every published optimizer-update adapter, including update zero (the initial, untrained adapter). Training is planned for 4 epochs on the same 2,048-question training split as the earlier STRIDE runs. The global…

GRPO with group-standardized answer correctness, without a STRIDE diversity bonus. Training explicitly renders enablethinking=False, recorded as thinkingmode: false in the scientific contract. The pinned tokenizer and chat template remain unchanged. Use the same explicit keyword at inference, particularly for Qwen3-1.7B, whose default template enables thinking. Every checkpoint in this repository requires explicit nonthinking provenance. This public repository retains every published optimizer-update adapter, including update zero (the initial, untrained adapter). Training is planned for 4 epochs on the same 2,048-question training split as the earlier STRIDE runs. The global prompt batch…

STRIDE with nonnegative local step-diversity credit on eligible reasoning tokens. Training explicitly renders enablethinking=False, recorded as thinkingmode: false in the scientific contract. The pinned tokenizer and chat template remain unchanged. Use the same explicit keyword at inference, particularly for Qwen3-1.7B, whose default template enables thinking. Every checkpoint in this repository requires explicit nonthinking provenance. This public repository retains every published optimizer-update adapter, including update zero (the initial, untrained adapter). Training is planned for 4 epochs on the same 2,048-question training split as the earlier STRIDE runs. The global prompt batch is…

STRIDE with nonnegative local step-diversity credit on eligible reasoning tokens. Training explicitly renders enablethinking=False, recorded as thinkingmode: false in the scientific contract. The pinned tokenizer and chat template remain unchanged. Use the same explicit keyword at inference, particularly for Qwen3-1.7B, whose default template enables thinking. Every checkpoint in this repository requires explicit nonthinking provenance. This public repository retains every published optimizer-update adapter, including update zero (the initial, untrained adapter). Training is planned for 4 epochs on the same 2,048-question training split as the earlier STRIDE runs. The global prompt batch is…

STRIDE with nonnegative local step-diversity credit on eligible reasoning tokens. Training explicitly renders enablethinking=False, recorded as thinkingmode: false in the scientific contract. The pinned tokenizer and chat template remain unchanged. Use the same explicit keyword at inference, particularly for Qwen3-1.7B, whose default template enables thinking. Every checkpoint in this repository requires explicit nonthinking provenance. This public repository retains every published optimizer-update adapter, including update zero (the initial, untrained adapter). Training is planned for 4 epochs on the same 2,048-question training split as the earlier STRIDE runs. The global prompt batch is…

STRIDE with nonnegative local step-diversity credit on eligible reasoning tokens. Training explicitly renders enablethinking=False, recorded as thinkingmode: false in the scientific contract. The pinned tokenizer and chat template remain unchanged. Use the same explicit keyword at inference, particularly for Qwen3-1.7B, whose default template enables thinking. Every checkpoint in this repository requires explicit nonthinking provenance. This public repository retains every published optimizer-update adapter, including update zero (the initial, untrained adapter). Training is planned for 4 epochs on the same 2,048-question training split as the earlier STRIDE runs. The global prompt batch is…

STRIDE ablation using one pooled reasoning embedding per eligible response. Training explicitly renders enablethinking=False, recorded as thinkingmode: false in the scientific contract. The pinned tokenizer and chat template remain unchanged. Use the same explicit keyword at inference, particularly for Qwen3-1.7B, whose default template enables thinking. Every checkpoint in this repository requires explicit nonthinking provenance. This public repository retains every published optimizer-update adapter, including update zero (the initial, untrained adapter). Training is planned for 4 epochs on the same 2,048-question training split as the earlier STRIDE runs. The global prompt batch is 64…

STRIDE ablation distributing each response diversity bonus uniformly over its eligible tokens. Training explicitly renders enablethinking=False, recorded as thinkingmode: false in the scientific contract. The pinned tokenizer and chat template remain unchanged. Use the same explicit keyword at inference, particularly for Qwen3-1.7B, whose default template enables thinking. Every checkpoint in this repository requires explicit nonthinking provenance. This public repository retains every published optimizer-update adapter, including update zero (the initial, untrained adapter). Training is planned for 4 epochs on the same 2,048-question training split as the earlier STRIDE runs. The global…

A Llama-3-8B-Instruct checkpoint compressed with SVD-LLM to 60.0% of dense parameters, then edited by 10 of 10 rounds of iterative parameter-neutral swap selected by the gapiter rule (up to 0.1% of dense parameters per round; the full run's budget is 1.0%). This is a research artifact from a study of how SVD compression damages safety behaviour and which component-selection rule best repairs it. It is one cell of a grid over selection rules and budgets; it is not a general-purpose chat model. This checkpoint exists to measure safety/utility trade-offs under compression. Several arms in the grid are deliberately safety-degraded relative to Llama-3-8B-Instruct: compression alone raises…

A Llama-3-8B-Instruct checkpoint compressed with SVD-LLM to 60.0% of dense parameters, then edited by 8 of 10 rounds of iterative parameter-neutral swap selected by the gapiter rule (up to 0.1% of dense parameters per round; the full run's budget is 1.0%). This is a research artifact from a study of how SVD compression damages safety behaviour and which component-selection rule best repairs it. It is one cell of a grid over selection rules and budgets; it is not a general-purpose chat model. This checkpoint exists to measure safety/utility trade-offs under compression. Several arms in the grid are deliberately safety-degraded relative to Llama-3-8B-Instruct: compression alone raises…

A Llama-3-8B-Instruct checkpoint compressed with SVD-LLM to 60.0% of dense parameters, then edited by 9 of 10 rounds of iterative parameter-neutral swap selected by the gapiter rule (up to 0.1% of dense parameters per round; the full run's budget is 1.0%). This is a research artifact from a study of how SVD compression damages safety behaviour and which component-selection rule best repairs it. It is one cell of a grid over selection rules and budgets; it is not a general-purpose chat model. This checkpoint exists to measure safety/utility trade-offs under compression. Several arms in the grid are deliberately safety-degraded relative to Llama-3-8B-Instruct: compression alone raises…

A Llama-3-8B-Instruct checkpoint compressed with SVD-LLM to 50.0% of dense parameters, then edited by 1 of 10 rounds of iterative parameter-neutral swap selected by the gapiter rule (up to 0.1% of dense parameters per round; the full run's budget is 1.0%). This is a research artifact from a study of how SVD compression damages safety behaviour and which component-selection rule best repairs it. It is one cell of a grid over selection rules and budgets; it is not a general-purpose chat model. This checkpoint exists to measure safety/utility trade-offs under compression. Several arms in the grid are deliberately safety-degraded relative to Llama-3-8B-Instruct: compression alone raises…

A Llama-3-8B-Instruct checkpoint compressed with SVD-LLM to 50.0% of dense parameters, then edited by 2 of 10 rounds of iterative parameter-neutral swap selected by the gapiter rule (up to 0.1% of dense parameters per round; the full run's budget is 1.0%). This is a research artifact from a study of how SVD compression damages safety behaviour and which component-selection rule best repairs it. It is one cell of a grid over selection rules and budgets; it is not a general-purpose chat model. This checkpoint exists to measure safety/utility trade-offs under compression. Several arms in the grid are deliberately safety-degraded relative to Llama-3-8B-Instruct: compression alone raises…

A Llama-3-8B-Instruct checkpoint compressed with SVD-LLM to 50.0% of dense parameters, then edited by 3 of 10 rounds of iterative parameter-neutral swap selected by the gapiter rule (up to 0.1% of dense parameters per round; the full run's budget is 1.0%). This is a research artifact from a study of how SVD compression damages safety behaviour and which component-selection rule best repairs it. It is one cell of a grid over selection rules and budgets; it is not a general-purpose chat model. This checkpoint exists to measure safety/utility trade-offs under compression. Several arms in the grid are deliberately safety-degraded relative to Llama-3-8B-Instruct: compression alone raises…

A Llama-3-8B-Instruct checkpoint compressed with SVD-LLM to 50.0% of dense parameters, then edited by 4 of 10 rounds of iterative parameter-neutral swap selected by the gapiter rule (up to 0.1% of dense parameters per round; the full run's budget is 1.0%). This is a research artifact from a study of how SVD compression damages safety behaviour and which component-selection rule best repairs it. It is one cell of a grid over selection rules and budgets; it is not a general-purpose chat model. This checkpoint exists to measure safety/utility trade-offs under compression. Several arms in the grid are deliberately safety-degraded relative to Llama-3-8B-Instruct: compression alone raises…

A Llama-3-8B-Instruct checkpoint compressed with SVD-LLM to 50.0% of dense parameters, then edited by 5 of 10 rounds of iterative parameter-neutral swap selected by the gapiter rule (up to 0.1% of dense parameters per round; the full run's budget is 1.0%). This is a research artifact from a study of how SVD compression damages safety behaviour and which component-selection rule best repairs it. It is one cell of a grid over selection rules and budgets; it is not a general-purpose chat model. This checkpoint exists to measure safety/utility trade-offs under compression. Several arms in the grid are deliberately safety-degraded relative to Llama-3-8B-Instruct: compression alone raises…

A Llama-3-8B-Instruct checkpoint compressed with SVD-LLM to 50.0% of dense parameters, then edited by 6 of 10 rounds of iterative parameter-neutral swap selected by the gapiter rule (up to 0.1% of dense parameters per round; the full run's budget is 1.0%). This is a research artifact from a study of how SVD compression damages safety behaviour and which component-selection rule best repairs it. It is one cell of a grid over selection rules and budgets; it is not a general-purpose chat model. This checkpoint exists to measure safety/utility trade-offs under compression. Several arms in the grid are deliberately safety-degraded relative to Llama-3-8B-Instruct: compression alone raises…

A Llama-3-8B-Instruct checkpoint compressed with SVD-LLM to 50.0% of dense parameters, then edited by 7 of 10 rounds of iterative parameter-neutral swap selected by the gapiter rule (up to 0.1% of dense parameters per round; the full run's budget is 1.0%). This is a research artifact from a study of how SVD compression damages safety behaviour and which component-selection rule best repairs it. It is one cell of a grid over selection rules and budgets; it is not a general-purpose chat model. This checkpoint exists to measure safety/utility trade-offs under compression. Several arms in the grid are deliberately safety-degraded relative to Llama-3-8B-Instruct: compression alone raises…

A Llama-3-8B-Instruct checkpoint compressed with SVD-LLM to 50.0% of dense parameters, then edited by 8 of 10 rounds of iterative parameter-neutral swap selected by the gapiter rule (up to 0.1% of dense parameters per round; the full run's budget is 1.0%). This is a research artifact from a study of how SVD compression damages safety behaviour and which component-selection rule best repairs it. It is one cell of a grid over selection rules and budgets; it is not a general-purpose chat model. This checkpoint exists to measure safety/utility trade-offs under compression. Several arms in the grid are deliberately safety-degraded relative to Llama-3-8B-Instruct: compression alone raises…

A Llama-3-8B-Instruct checkpoint compressed with SVD-LLM to 50.0% of dense parameters, then edited by 10 of 10 rounds of iterative parameter-neutral swap selected by the gapiter rule (up to 0.1% of dense parameters per round; the full run's budget is 1.0%). This is a research artifact from a study of how SVD compression damages safety behaviour and which component-selection rule best repairs it. It is one cell of a grid over selection rules and budgets; it is not a general-purpose chat model. This checkpoint exists to measure safety/utility trade-offs under compression. Several arms in the grid are deliberately safety-degraded relative to Llama-3-8B-Instruct: compression alone raises…

meta-llama/Llama-2-7b-chat-hf compressed with Basis Sharing using the authors' code then recovered by LoRA on the coefficients alone. (the shared-basis fit) -> LoRA on the coefficients, bases frozen -> merge C' = C + (alpha/r)BA -> fold W = C' @ B to dense 2048 tokens, seed 42 (upstream hard-codes seed 2023; this project calibrates every method at one seed) 2 epochs, lr 0.0001, batch 64, on yahma/alpaca-cleaned Only the coefficients were trained. The shared and per-layer bases are frozen and bit-identical to the compressed model, so every weight keeps rank <= k, each group still shares one basis, and the parameter budget survives recovery exactly. This is not Basis Sharing's own LoRA…