---
license: other
base_model: Nanbeige/Nanbeige4.2-3B
library_name: transformers.js
pipeline_tag: text-generation
tags:
- onnx
- lora
- toolace
- nanbeige
- peft
- webgpu
---

# Nanbeige4.2-3B ToolACE LoRA (n3000) — ONNX

Merged LoRA fine-tune of [`Nanbeige/Nanbeige4.2-3B`](https://huggingface.co/Nanbeige/Nanbeige4.2-3B) on
[Team-ACE/ToolACE](https://huggingface.co/datasets/Team-ACE/ToolACE)
(n3000 subset). Prefill-only ONNX with **num_loops=2 unrolled**.

## Files
| Path | Notes |
|------|-------|
| `onnx/model.onnx` | Graph (opset 18) |
| `onnx/model.onnx_data` | Consolidated FP16 weights |
| tokenizer / config | From base + FT |

## IO
- **input**: `input_ids` [batch, seq]
- **output**: `logits` [batch, seq, vocab]
- Prefill (`use_cache=False`). Looped Transformer unrolled: 44 layer calls + 2 norms.

## Validation
- Unrolled vs stock max abs diff: `0.000e+00`
- ORT vs torch max abs diff: `5.273e-02` (FP16 tol 5e-2)

## Train recipe
- LoRA r=16, attn+MLP (no lm_head)
- TRL SFTTrainer, completion_only_loss
- batch 1 × accum 16, seq 2048, 1 epoch, transformers 4.51
- Adapter: `nanbeige_toolace_r16_n3000`

## Runtime note
Nanbeige generation on **transformers 5.x is broken** (DynamicCache + loop KV).
Use **transformers==4.51.0** for HF/PyTorch inference. ONNX path is independent.