GLM-5.3 is live. Run Z.ai's latest model on Friendli Model APIs. Try it today ➜
Congo-digital-service
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
Congo-digital-service
Model Tree
Qwen/Qwen2-VL-2B-Instruct
Input Modalities
Output Modalities
Supported Functionality
| Task | Transcription of Lingala text in images (OCR) |
| Base model | Qwen/Qwen2-VL-2B-Instruct |
| Adapters | LoRA rank 32, alpha 64, dropout 0.1 |
| Test result | CER 11.3 %, i.e. 88.7 % of characters correct (normalised, micro, 50 images) — details on the merged model's card |
| Licence | Apache 2.0 |
| Element | Detail |
|---|---|
| LoRA targets | q_proj, k_proj, v_proj, o_proj, gate_proj, up_proj, down_proj, merger.mlp.0, merger.mlp.2 |
| Fully saved modules | embed_tokens, lm_head |
| Added tokens | (in ) |
python
import torchfrom PIL import Imagefrom huggingface_hub import hf_hub_downloadfrom peft import PeftModelfrom safetensors import safe_openfrom transformers import Qwen2VLForConditionalGeneration, AutoProcessor base_id = "Qwen/Qwen2-VL-2B-Instruct"adapter_id = "Congo-digital-service/qwen-vl-lingala-qlora-adaptaters-vf" processor = AutoProcessor.from_pretrained(adapter_id) # extended tokenizermodel = Qwen2VLForConditionalGeneration.from_pretrained( base_id, torch_dtype=torch.bfloat16, device_map="auto") # The adapter stores the full embedding matrix: match its number of rows.path = hf_hub_download(adapter_id, "adapter_model.safetensors")with safe_open(path, "pt") as f: key = next(k for k in f.keys() if "embed_tokens" in k) n_rows = f.get_slice(key).get_shape()[0]if model.get_input_embeddings().weight.shape[0] != n_rows: model.resize_token_embeddings(n_rows) model = PeftModel.from_pretrained(model, adapter_id).eval() image = Image.open("line.png").convert("RGB")messages = [{"role": "user", "content": [ {"type": "image"}, {"type": "text", "text": "Transcris le texte de cette image."},]}]prompt = processor.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)inputs = processor(text=[prompt], images=[image], return_tensors="pt").to(model.device)with torch.no_grad(): out = model.generate(**inputs, max_new_tokens=256, do_sample=False)print(processor.batch_decode(out[:, inputs["input_ids"].shape[1]:], skip_special_tokens=True)[0])Keep max_new_tokens bounded: about 3 outputs in 50 loop until the length limit. For most uses, the merged model is simpler and gives the same outputs.
417 image–text examples (278 lines, ɔ/ɛ lines oversampled ×2), 5 epochs, learning rate 1 × 10⁻⁴, batch 1 × 8, linear schedule with 10 % warm-up, PagedAdamW 8-bit, Google Colab, August 2026. The epoch-1 checkpoint was retained; this choice was made on the test set. Full details, per-epoch scores, evaluation and limitations: see the merged model's card.
Qwen/Qwen2-VL-2B-Instruct.Congo Digital Services (CDS SARL), https://www.congo-digital.com/, with the MALOBA community, SIL Congo, the Ministry of Posts, Telecommunications and the Digital Economy of the Republic of Congo and UNDP Congo.
Created in August 2026. Card updated on 30 September 2026.
ƆƐɔɛadded_tokens.json| Tokenizer and processor | included: load the processor from this repository, not from the base model, or the four added tokens are missing |
| Size | about 5.7 GB: the embedding and output matrices are saved in full, because the vocabulary was extended |