laya-flash (LFM2.5 baseline)

This is a format baseline, not a usable model yet. It publishes the LiquidAI/LFM2.5-Encoder-230M backbone inside a laya-flash checkpoint, with the decision head freshly initialized and untrained. Expect near-uniform probabilities and near-zero confidence on every question — that is the honest behaviour of this checkpoint, not a bug. Fine-tuned versions will replace these files in place.

What is in the checkpoint

Part	State
Encoder (LFM2.5-Encoder-230M, bidirectional patches)	LiquidAI pretrained weights, unchanged
Decision head (transformer layers, scorer, act head)	Random initialization (seed 42)
Temperatures	Neutral (1.0 — nothing was calibrated)

The config (rl_agent_config.json) mirrors the shipped laya-flash checkpoints (act_costs, cost_wrong_act, max_prefixes, amp_dtype), so a fine-tune can start from this checkpoint with the head shapes it expects.

Usage

pip install 'laya-flash[lfm2]'   # LFM2 needs transformers >= 4.55

import laya_flash
agent = laya_flash.Agent("nampham1106/laya-flash")
res = agent.system_one("Customer was charged twice and wants money back.", {
    "intent": {"type": "choice", "instructions": "What does the customer want?",
               "criteria": {"refund": "money back", "tech": "a bug", "other": "anything else"}},
})

The checkpoint layout is the standard laya-flash one: rl_agent_config.json, model.safetensors (full state dict, fp32), tokenizer/, encoder/config.json.

License and attribution

The encoder weights are Liquid AI, Inc.'s, distributed under the LFM Open License v1.0 (see LICENSE). This checkpoint is a Derivative Work and ships under the same license. In short: commercial use is licensed below $10M annual revenue; redistribution must include the license and attribution. laya-flash's bidirectional LFM2 patches are adapted from LiquidAI's Apache-2.0 modeling_lfm2_bidirectional.py.