Upload README.md with huggingface_hub
Browse files
README.md
ADDED
|
@@ -0,0 +1,59 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
---
|
| 2 |
+
license: apache-2.0
|
| 3 |
+
language:
|
| 4 |
+
- es
|
| 5 |
+
base_model: CodeStrux-Tech/tac-1
|
| 6 |
+
base_model_relation: quantized
|
| 7 |
+
pipeline_tag: text-generation
|
| 8 |
+
tags:
|
| 9 |
+
- fp8
|
| 10 |
+
- compressed-tensors
|
| 11 |
+
- llmcompressor
|
| 12 |
+
- address-parsing
|
| 13 |
+
- costa-rica
|
| 14 |
+
- qwen3
|
| 15 |
+
---
|
| 16 |
+
|
| 17 |
+
# tac-1-fp8-dynamic — FP8 dynamic quantization of tac-1
|
| 18 |
+
|
| 19 |
+
## Overview
|
| 20 |
+
|
| 21 |
+
tac-1-fp8-dynamic is quantized from [`CodeStrux-Tech/tac-1`](https://huggingface.co/CodeStrux-Tech/tac-1) using llmcompressor 0.12.0 with the FP8_DYNAMIC scheme. Targets are `Linear` layers; `lm_head` is ignored. The result is a single 4.2 GB shard.
|
| 22 |
+
|
| 23 |
+
sha256: `e0eb5c3c62e8a472eaca6f86064efb2597863f0688ac3ca1fd9a325390d9c30f`
|
| 24 |
+
|
| 25 |
+
## Serving
|
| 26 |
+
|
| 27 |
+
```bash
|
| 28 |
+
vllm serve CodeStrux-Tech/tac-1-fp8-dynamic --served-model-name tac-1 --max-model-len 4096
|
| 29 |
+
# On hosts without nvcc, disable the FlashInfer sampler JIT:
|
| 30 |
+
# VLLM_USE_FLASHINFER_SAMPLER=0 vllm serve ...
|
| 31 |
+
```
|
| 32 |
+
|
| 33 |
+
No `--quantization` flag is needed — vLLM auto-detects the FP8 format from the model config. The FP8 model is a drop-in replacement for the bf16 original; the client snippet from [`CodeStrux-Tech/tac-1`](https://huggingface.co/CodeStrux-Tech/tac-1) works unchanged:
|
| 34 |
+
|
| 35 |
+
```python
|
| 36 |
+
from tico.extractor.clients.vllm import VLLMExtractor
|
| 37 |
+
|
| 38 |
+
extractor = VLLMExtractor() # VLLM_BASE_URL=http://localhost:8000/v1, TICO_VLLM_MODEL=tac-1
|
| 39 |
+
addr = extractor.extract("del antiguo higuerón de San Pedro, 100 metros sur")
|
| 40 |
+
print(addr.model_dump_json(indent=2))
|
| 41 |
+
```
|
| 42 |
+
|
| 43 |
+
## Evaluation
|
| 44 |
+
|
| 45 |
+
| metric | tac-1 (bf16) | tac-1 (FP8) | teacher gpt-4.1 |
|
| 46 |
+
|---|---|---|---|
|
| 47 |
+
| parse rate | 1.000 | 1.000 | 1.000 |
|
| 48 |
+
| anchor_name F1 | 0.979 | 0.979 | 0.872 |
|
| 49 |
+
| direction F1 | 0.995 | 0.984 | 1.000 |
|
| 50 |
+
| distance F1 | 0.995 | 0.984 | 1.000 |
|
| 51 |
+
| perfect-anchor ≤250 m | 93.6% | 93.6% | 93.6% |
|
| 52 |
+
| perfect-anchor median | 0.0 m | 0.0 m | 0.0 m |
|
| 53 |
+
| gazetteer ≤250 m | 80.9% | 80.9% | 83.0% |
|
| 54 |
+
|
| 55 |
+
## Training data attribution
|
| 56 |
+
|
| 57 |
+
Contains information from OpenStreetMap (https://www.openstreetmap.org/copyright), which is made available under the Open Database License (ODbL) 1.0. © OpenStreetMap contributors.
|
| 58 |
+
|
| 59 |
+
For full training details, architecture, evaluation, and limitations, see [`CodeStrux-Tech/tac-1`](https://huggingface.co/CodeStrux-Tech/tac-1).
|