A drop-in, numerically identical variant of LiquidAI/LFM2.5-1.2B-Instruct whose linear sub-modules are renamed to the Llama convention, so LoRA tooling that defaults to o_proj / gate_proj / up_proj / down_proj targets the full attention + MLP surface instead of q/k/v_proj only.

This is not a new model. The weights are bit-for-bit those of the base; only the attribute and tensor names change. Verified numerically identical to the base on the same input (max absolute logit difference = 0.0).

LFM2 names its attention output out_proj and its SwiGLU MLP w1/w3/w2. A LoRA config that lists the standard Llama names matches only q/k/v_proj and silently skips the attention output and the MLP. This variant renames those modules so the same default trains the whole linear surface.

Conv blocks (conv.in_proj, conv.out_proj) are unchanged.

The renamed attributes come from the bundled modeling_lfm2_ertas.py, so loading needs trust_remote_code=True:

For inference or GGUF export, use the stock base and map adapter names back (o_proj to out_proj, gate/up/down_proj to w1/w3/w2).

Derived from LiquidAI/LFM2.5-1.2B-Instruct, copyright Liquid AI, distributed under the LFM Open License v1.0 (see LICENSE). This naming variant was prepared by Ertas AI for internal LoRA fine-tuning tooling. All model capabilities, weights, and credit belong to Liquid AI.