GLM-5.3 is live. Run Z.ai's latest model on Friendli Model APIs. Try it today ➜
DevTheusP
Available on FriendliAI
Run this model inference on single tenant GPU with unmatched speed and reliability at scale.
Model Details
Model Provider
DevTheusP
Model Tree
Qwen/Qwen2.5-0.5B-Instruct
Input Modalities
Output Modalities
Supported Functionality
| caminho | o que é |
|---|---|
| raiz | modelo merged (fp16, safetensors), pronto para transformers ou vLLM |
adapter/ | só o adaptador LoRA, para aplicar sobre o Qwen/Qwen2.5-0.5B-Instruct |
gguf/ | GGUF q4_K_M e Modelfile para Ollama ou llama.cpp |
train_log.json | hiperparâmetros e curva de perda |
O modelo foi treinado com um prompt fixo; use exatamente o SYSTEM abaixo (o mesmo do benchmark)
e a mensagem de usuário no formato Chaves existentes: / Mensagem:.
python
import jsonfrom transformers import AutoModelForCausalLM, AutoTokenizer repo = "DevTheusP/qwen2.5-0.5b-fact-extractor-ptbr"tok = AutoTokenizer.from_pretrained(repo)model = AutoModelForCausalLM.from_pretrained(repo) from huggingface_hub import hf_hub_downloadSYSTEM = open(hf_hub_download(repo, "system_prompt.txt")).read() # o mesmo prompt do benchmarkuser = "Chaves existentes:\n- billing.database_engine\n\nMensagem:\nDecidimos trocar o banco do faturamento: sai MySQL, entra PostgreSQL."msgs = [{"role": "system", "content": SYSTEM}, {"role": "user", "content": user}]ids = tok.apply_chat_template(msgs, add_generation_prompt=True, return_tensors="pt")out = model.generate(ids, max_new_tokens=256, do_sample=False)print(json.loads(tok.decode(out[0, ids.shape[1]:], skip_special_tokens=True)))# {"facts": [{"subject_key": "billing.database_engine", "value": "PostgreSQL", "kind": "semantic", ...}]}Com Ollama (pasta gguf/): ollama create fact-extractor-0.5b -f Modelfile. Para saída sempre
válida, passe o JSON Schema no campo format (Ollama) ou response_format (vLLM), como o benchmark faz.
Medidos no notebook do Kaggle. Referência = rótulos do professor, salvo indicação.
| modelo | conjunto | modo | schema_valid | strict_f1 | value_f1 | reuse_accuracy | kind_accuracy | empty_accuracy | false_empty_rate | latency_p50_s | tokens_per_s |
|---|---|---|---|---|---|---|---|---|---|---|---|
| aluno-base | test_bench | ref=professor | guided | 100.0% | 1.7% | 7.5% | 100.0% | 11.1% | 0.0% | 0.0% | 0.72 |
| aluno-destilado | test_bench | ref=professor | guided | 100.0% | 28.6% | 76.2% | 75.0% | 75.0% | 98.3% | 27.3% | 0.06 |
| professor | test_bench | ref=professor | guided | 100.0% | 90.9% | 100.0% | 100.0% | 100.0% | 100.0% | 0.0% | 0.26 |
| aluno-base | test_synth | ref=professor | free | 0.0% | - | - | 0.0% | - | 0.0% | 0.0% | 0.30 |
| aluno-destilado | test_synth | ref=professor | free | 99.7% | 25.9% | 61.9% | 53.1% | 61.5% | 80.0% | 13.3% | 0.28 |
| professor | test_synth | ref=professor | free | 0.0% | - | - | 0.0% | - | 0.0% | 0.0% | 1.30 |
| extrator | memória | retrieval | ret_update | ret_scope | ret_rare_fact | ret_injection | falhas de JSON |
|---|---|---|---|---|---|---|---|
| aluno-base | v1_naive | 62.0% | 100.0% | 100.0% | 33.7% | 0.0% | 58 |
| aluno-base | v2_keyed | 83.3% | 100.0% | 100.0% | 33.1% | 100.0% | 58 |
| aluno-base | v3_decay_rank | 53.4% | 99.4% | 12.8% | 1.5% | 100.0% | 58 |
| aluno-base | v4_decay_tier | 53.1% | 100.0% | 8.6% | 3.8% | 100.0% | 58 |
| aluno-base | v5_guarded | 97.6% | 99.0% | 100.0% | 91.4% | 100.0% | 58 |
| aluno-destilado | v1_naive | 62.0% | 100.0% | 100.0% | 33.7% | 0.0% | 0 |
| aluno-destilado | v2_keyed | 100.0% | 100.0% | 100.0% | 100.0% | 100.0% | 0 |
| aluno-destilado | v3_decay_rank | 71.7% | 96.4% | 54.4% | 36.6% | 100.0% | 0 |
| aluno-destilado | v4_decay_tier | 69.4% | 100.0% | 37.5% | 40.0% | 100.0% | 0 |
| aluno-destilado | v5_guarded | 100.0% | 100.0% | 100.0% | 100.0% | 100.0% | 0 |
| professor | v1_naive | 62.0% | 100.0% | 100.0% | 33.7% | 0.0% | 0 |
| professor | v2_keyed | 83.4% | 100.0% | 100.0% | 33.7% | 100.0% | 0 |
| professor | v3_decay_rank | 54.7% | 97.5% | 14.1% | 7.4% | 100.0% | 0 |
| professor | v4_decay_tier | 55.2% | 100.0% | 8.6% | 12.1% | 100.0% | 0 |
| professor | v5_guarded | 97.9% | 100.0% | 100.0% | 91.4% | 100.0% | 0 |
| publicado (Ollama, qwen2.5:7b) | v2_keyed | 98% | |||||
| publicado (Ollama, qwen2.5:7b) | v1_naive | 62% |
SYSTEM ou outro formato de mensagem, o comportamento não foi avaliado.Apache-2.0, a mesma do modelo base Qwen2.5-0.5B-Instruct. Os rótulos de treino vêm do Qwen2.5-7B-Instruct, também Apache-2.0.
import jsonfrom transformers import AutoModelForCausalLM, AutoTokenizer repo = "DevTheusP/qwen2.5-0.5b-fact-extractor-ptbr"tok = AutoTokenizer.from_pretrained(repo)model = AutoModelForCausalLM.from_pretrained(repo) from huggingface_hub import hf_hub_downloadSYSTEM = open(hf_hub_download(repo, "system_prompt.txt")).read() # o mesmo prompt do benchmarkuser = "Chaves existentes:\n- billing.database_engine\n\nMensagem:\nDecidimos trocar o banco do faturamento: sai MySQL, entra PostgreSQL."msgs = [{"role": "system", "content": SYSTEM}, {"role": "user", "content": user}]ids = tok.apply_chat_template(msgs, add_generation_prompt=True, return_tensors="pt")out = model.generate(ids, max_new_tokens=256, do_sample=False)print(json.loads(tok.decode(out[0, ids.shape[1]:], skip_special_tokens=True)))# {"facts": [{"subject_key": "billing.database_engine", "value": "PostgreSQL", "kind": "semantic", ...}]}