GLM-5.2 is live. #1 throughput on OpenRouter, pay-per-token on FriendliAI. Try it today ➜
LockNToad
Available on FriendliAI
Run this model inference on single tenant GPU with unmatched speed and reliability at scale.
Run this model inference with full control and performance in your environment.
Model Details
Model Provider
LockNToad
Model Tree
Qwen/Qwen2.5-VL-7B-Instruct
Input Modalities
Output Modalities
Supported Functionality
cartolegend-7b-r16-rc3f-preserve-rc3e-mfs-lr1e8-steps32-128k-accum8.Qwen/Qwen2.5-VL-7B-Instruct.| Split | Matched / Gold | Precision | 32px | 64px | Empty-Gold FP |
|---|---|---|---|---|---|
| public-hard regression | 20 / 20 | 1.000 | 0.850 | 0.950 | 0 |
| B961 diagnostic | 2 / 2 | 1.000 | 1.000 | 1.000 | 0 |
| eval5 reviewed | 25 / 25 | 1.000 | 0.920 | 1.000 | 0 |
| eval15 reviewed | 60 / 64 | 1.000 | 0.900 | 0.967 | 0 |
| Unseen5 reviewed | 9 / 9 | 1.000 | 0.222 | 0.778 | 0 |
| post60 reviewed | 43 / 47 | 0.977 | 0.744 | 0.930 | 0 |
| locked public blind | 41 / 106 | 0.394 | 0.000 | 0.000 | 0 |
Six-split recall is 159/167 (0.9520958), which clears the RC3 objective of >0.95. The locked public blind split improves over the RC2 packaged baseline of 39/106.
Public-hard is train-seen regression evidence only and must not be presented as held-out quality evidence.
adapter_model.safetensors and adapter_config.json: PEFT LoRA adapter and public base-model config.schema/cartolegend_schema.json: scoped point-symbol output schema.eval/package_guard/: copied package-guard metrics, final predictions, and per-split summaries.RC3_PACKAGE_GUARD_VERIFY.*: candidate-specific gate verifier from the per-split package artifacts.TRAIN_VERIFY_RC3F.*: RC3f train/source-disjoint verifier.VERIFY_RC3_RELEASE_RESULT.*: release-bundle verifier output.scripts/: deterministic pipeline and verification scripts.Use for research and human-audited extraction of visible point-symbol entries from geologic or resource-map legend crops. The output should be reviewed before downstream use.
Not intended for autonomous geologic interpretation, legal boundary decisions, navigation, safety-critical mapping, or broad non-point legend extraction.
python
from transformers import AutoProcessor, Qwen2_5_VLForConditionalGenerationfrom peft import PeftModelbase_id = "Qwen/Qwen2.5-VL-7B-Instruct"adapter_id = "path/to/cartolegend-qwen2p5vl-7b-lora-rc3-package-guard"processor = AutoProcessor.from_pretrained(base_id, use_fast=True)model = Qwen2_5_VLForConditionalGeneration.from_pretrained(base_id,torch_dtype="auto",device_map="auto",)model = PeftModel.from_pretrained(model, adapter_id)
For the reported metrics, run the packaged pipeline scripts and verifier rather than raw model generation alone.
bash
python3 scripts/verify_cartolegend_rc3_release_bundle.py \--release-dir release/cartolegend-qwen2p5vl-7b-lora-rc3-package-guard \--require-pass