experimentalmachines/LFM2.5-1.2B-Instruct-heretic

Abliterated ("uncensored") version of LiquidAI/LFM2.5-1.2B-Instruct, produced with heretic: directional ablation of the refusal direction in the residual stream, with a TPE search over per-layer ablation weights that minimises both the refusal rate and the KL divergence from the original model. Weights are merged, so this loads as a plain Lfm2ForCausalLM. Chat template, tool-calling tokens and everything else are unchanged from the base model.

heretic Pareto front for this run

study: ./checkpoints/LiquidAI--LFM2--5-1--2B-Instruct.jsonl
completed trials: 200 / 200

idx trial refusals       KL  direction_index
  0   133     0.06   0.0527  10.198270144464333
  1   150     0.07   0.0402  10.32046799180982
  2   157     0.11   0.0344  10.352822192960996
  3   170     0.17   0.0316  10.690289415212908
  4   169     0.18   0.0283  10.969840642119093
  5   195     0.41   0.0232  9.737856473201463
  6    63     0.43   0.0227  None
  7   143     0.46   0.0217  11.294404725756008
  8   158     0.49   0.0174  11.119204885113456
  9   186     0.53   0.0170  10.5853942209232
 10   191     0.63   0.0157  11.178660347171663
 11    72     0.66   0.0124  11.153256538028568
 12   161     0.79   0.0102  10.51671121861693
 13   102     0.84   0.0097  11.205828382998588
 14    59     0.87   0.0093  10.78726439886388
 15    76     0.90   0.0061  None
 16   164     0.91   0.0058  10.51819047303982
 17   182     0.96   0.0028  None
 18    26     0.98   0.0006  None
selected pareto index: 0

ExecuTorch (CPU, KleidiAI on Arm)

file	max context (tokens)	size
executorch/lfm2_5_1_2b_heretic_8da4w_32k.pte	32,768	0.83 GB

Each is an ExecuTorch 1.4.1 program for the XNNPACK backend built with ExecuTorch's own LFM2 recipe: every linear layer is quantised to 8-bit dynamic per-token activations and 4-bit grouped weights with group size 32 (8da4w, the layout Arm's KleidiAI int4 kernels are written for), int8 embedding table, static KV cache, fused SDPA, 2048-token prefill chunk (longer prompts are prefilled in chunks; the context window is the figure in the table). XNNPACK dispatches these matmuls to KleidiAI on every Arm CPU (Apple Silicon, Snapdragon, Dimensity, Tensor, Graviton) in the stock ExecuTorch build; on x86 it falls back to AVX kernels. The program is the same file on every platform.

The KV cache for the full window is allocated when the program loads (fp32, about 0.8 GB for 1.2B and 1.1 GB for 2.6B); the runtime attends over filled positions only, so the window costs memory, not speed. Run it with llama_main from ExecuTorch (examples/models/llama) or the python runner, with this repo's tokenizer.json and the prompt format <|startoftext|><|im_start|>user\n...<|im_end|>\n<|im_start|>assistant\n.

Pipeline

Reproducible end to end from https://github.com/ (see the repository this was built with): make setup abliterate pareto, make save TRIAL=<idx>, make export, make publish.

Use responsibly; the safety training of the base model has been removed on purpose.