GLM-5.3 is live. Run Z.ai's latest model on Friendli Model APIs. Try it today ➜
Hedmm
Available on FriendliAI
Run this model inference on single tenant GPU with unmatched speed and reliability at scale.
Model Details
Model Provider
Hedmm
Model Tree
Qwen/Qwen2.5-1.5B-Instruct
Input Modalities
Output Modalities
Supported Functionality
| Model | Accuracy (95% CI) | Macro-F1 |
|---|---|---|
| Base Qwen2.5-1.5B-Instruct, same prompt | 36.3% (34.6–38.1) | 0.339 |
| This adapter (LoRA) | 92.1% (91.1–93.1) | 0.921 |
eval/preds.jsonl, together with metrics.json and train_log.jsonl.python
import jsonimport torchfrom huggingface_hub import hf_hub_downloadfrom peft import PeftModelfrom transformers import AutoModelForCausalLM, AutoTokenizer base_id = "Qwen/Qwen2.5-1.5B-Instruct"adapter_id = "Hedmm/qwen2.5-1.5b-instruct-banking77-lora" tokenizer = AutoTokenizer.from_pretrained(adapter_id)model = AutoModelForCausalLM.from_pretrained(base_id, dtype=torch.bfloat16, device_map="auto")model = PeftModel.from_pretrained(model, adapter_id).eval()labels = json.load(open(hf_hub_download(adapter_id, "labels.json"))) def classify(message: str) -> str: """Return the predicted Banking77 intent for one customer message.""" chat = [ {"role": "system", "content": "You are an intent classifier for a bank's customer support."}, {"role": "user", "content": "Classify the customer message into exactly one of these intents:\n" + ", ".join(labels) + f"\n\nMessage: {message}\nAnswer with the intent label only."}, ] prompt = tokenizer.apply_chat_template(chat, tokenize=False, add_generation_prompt=True) inputs = tokenizer(prompt, return_tensors="pt", add_special_tokens=False).to(model.device) output = model.generate(**inputs, max_new_tokens=16, do_sample=False) answer = tokenizer.decode(output[0, inputs["input_ids"].shape[1]:], skip_special_tokens=True) return answer.strip().split("\n")[0].strip().strip(".") print(classify("I still haven't received my new card, it's been two weeks."))Use exactly this prompt, with the label list in the order of labels.json (sorted alphabetically). The adapter was trained on it.
| Setting | Value |
|---|---|
| Base model | Qwen/Qwen2.5-1.5B-Instruct, bf16 |
| Data | Banking77 train split, 10,003 messages; loss only on the answer tokens |
| LoRA | rank 16, alpha 32, dropout 0.05, on q, k, v, o, gate, up and down projections (18.5 M trainable parameters, 1.2% of the model) |
| Optimisation | AdamW, learning rate 2e-4, 3% warm-up then cosine decay, batch 8 × 4 gradient accumulation, 2 epochs, seed 42 |
| Hardware | one NVIDIA H100 PCIe: 16.6 min of training, 10.6 GB peak GPU memory |
import jsonimport torchfrom huggingface_hub import hf_hub_downloadfrom peft import PeftModelfrom transformers import AutoModelForCausalLM, AutoTokenizer base_id = "Qwen/Qwen2.5-1.5B-Instruct"adapter_id = "Hedmm/qwen2.5-1.5b-instruct-banking77-lora" tokenizer = AutoTokenizer.from_pretrained(adapter_id)model = AutoModelForCausalLM.from_pretrained(base_id, dtype=torch.bfloat16, device_map="auto")model = PeftModel.from_pretrained(model, adapter_id).eval()labels = json.load(open(hf_hub_download(adapter_id, "labels.json"))) def classify(message: str) -> str: """Return the predicted Banking77 intent for one customer message.""" chat = [ {"role": "system", "content": "You are an intent classifier for a bank's customer support."}, {"role": "user", "content": "Classify the customer message into exactly one of these intents:\n" + ", ".join(labels) + f"\n\nMessage: {message}\nAnswer with the intent label only."}, ] prompt = tokenizer.apply_chat_template(chat, tokenize=False, add_generation_prompt=True) inputs = tokenizer(prompt, return_tensors="pt", add_special_tokens=False).to(model.device) output = model.generate(**inputs, max_new_tokens=16, do_sample=False) answer = tokenizer.decode(output[0, inputs["input_ids"].shape[1]:], skip_special_tokens=True) return answer.strip().split("\n")[0].strip().strip(".") print(classify("I still haven't received my new card, it's been two weeks."))