Upload README.md with huggingface_hub
Browse files
README.md
ADDED
|
@@ -0,0 +1,44 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
---
|
| 2 |
+
license: other
|
| 3 |
+
license_name: lfm1.0
|
| 4 |
+
base_model: LiquidAI/LFM2.5-2.6B
|
| 5 |
+
library_name: transformers.js
|
| 6 |
+
pipeline_tag: text-generation
|
| 7 |
+
tags:
|
| 8 |
+
- onnx
|
| 9 |
+
- lora
|
| 10 |
+
- toolace
|
| 11 |
+
- lfm2.5
|
| 12 |
+
- peft
|
| 13 |
+
- webgpu
|
| 14 |
+
---
|
| 15 |
+
|
| 16 |
+
# LFM2.5-2.6B ToolACE LoRA (n=n3000) — ONNX
|
| 17 |
+
|
| 18 |
+
Merged LoRA fine-tune of [`LiquidAI/LFM2.5-2.6B`](https://huggingface.co/LiquidAI/LFM2.5-2.6B) on
|
| 19 |
+
[Team-ACE/ToolACE](https://huggingface.co/datasets/Team-ACE/ToolACE)
|
| 20 |
+
(n3000 subset), exported for **in-browser ORT Web**.
|
| 21 |
+
|
| 22 |
+
## Files
|
| 23 |
+
| Path | Notes |
|
| 24 |
+
|------|-------|
|
| 25 |
+
| `onnx/model.onnx` | Graph (opset 18) |
|
| 26 |
+
| `onnx/model.onnx_data` | Consolidated FP32 weights |
|
| 27 |
+
| tokenizer / config | From base + FT |
|
| 28 |
+
|
| 29 |
+
## IO
|
| 30 |
+
- **inputs**: `input_ids` [batch, seq], `attention_mask` [batch, seq]
|
| 31 |
+
- **outputs**: `logits` [batch, seq, vocab]
|
| 32 |
+
- Prefill-style (`use_cache=False`). No KV-cache I/O.
|
| 33 |
+
|
| 34 |
+
## Validation
|
| 35 |
+
FP32 ORT vs PyTorch max abs diff: `1.335e-04` (threshold 1e-3).
|
| 36 |
+
|
| 37 |
+
## Train recipe
|
| 38 |
+
- LoRA r=16, attn+MLP (no lm_head)
|
| 39 |
+
- TRL SFTTrainer, completion_only_loss
|
| 40 |
+
- batch 1 × accum 16, seq 2048, 1 epoch
|
| 41 |
+
- Adapter path: `lfm25_toolace_r16_n3000`
|
| 42 |
+
|
| 43 |
+
## License
|
| 44 |
+
LFM 1.0 (inherited from base). ToolACE dataset license applies to fine-tune data.
|