nicolasembleton/LFM2.5-2.6B-ToolACE-n3000-ONNX

LFM2.5-2.6B ToolACE LoRA (n=n3000) — ONNX

Merged LoRA fine-tune of `LiquidAI/LFM2.5-2.6B` on Team-ACE/ToolACE (n3000 subset), exported for in-browser ORT Web.

Files

IO

• —inputs: input_ids [batch, seq], attention_mask [batch, seq]
• —outputs: logits [batch, seq, vocab]
• —Prefill-style (use_cache=False). No KV-cache I/O.

Validation

FP32 ORT vs PyTorch max abs diff: 1.335e-04 (threshold 1e-3).

Train recipe

• —LoRA r=16, attn+MLP (no lm_head)
• —TRL SFTTrainer, completiononlyloss
• —batch 1 × accum 16, seq 2048, 1 epoch
• —Adapter path: lfm25_toolace_r16_n3000

License

LFM 1.0 (inherited from base). ToolACE dataset license applies to fine-tune data.