GLM-5.3 is live. Run Z.ai's latest model on Friendli Model APIs. Try it today ➜
Eternity5551
Available on FriendliAI
Run this model inference on single tenant GPU with unmatched speed and reliability at scale.
Model Details
Model Provider
Eternity5551
Model Tree
Qwen/Qwen3-1.7B-Base
Input Modalities
Output Modalities
Supported Functionality
1e-4. Loss is computed only on the code completion.0.14723. The adapter was selected by validation loss, not benchmark score.| Benchmark | Base | Full SFT | This LoRA SFT adapter |
|---|---|---|---|
| HumanEval+ v0.1.10 | 31/164 (18.9%) | 67/164 (40.9%) | 82/164 (50.0%) |
| MBPP+ v0.2.0 | 214/378 (56.6%) | 229/378 (60.6%) | 237/378 (62.7%) |
Strict pass@1 requires both original and Plus tests to pass. The same frozen tasks, raw prompt, greedy decoding, 512-token cap and restricted Docker judge were used for all three models. The different SFT learning rates and optimizers mean this result should not be attributed solely to the LoRA update mechanism. Per-task results and evaluation details are public.
python
from peft import PeftModelfrom transformers import AutoModelForCausalLM, AutoTokenizer base = AutoModelForCausalLM.from_pretrained( "Qwen/Qwen3-1.7B-Base", revision="ea980cb0a6c2ae4b936e82123acc929f1cec04c1", dtype="auto", device_map="auto",)adapter = "Eternity5551/Qwen3-1.7B-Python-Code-LoRA-SFT"model = PeftModel.from_pretrained(base, adapter).eval()tokenizer = AutoTokenizer.from_pretrained(adapter) prompt = '"""\nWrite a Python function double_even(nums) that doubles only even integers.\n"""\n\n'inputs = tokenizer(prompt, return_tensors="pt").to(model.device)output = model.generate(**inputs, do_sample=False, max_new_tokens=512, pad_token_id=tokenizer.eos_token_id)print(tokenizer.decode(output[0, inputs.input_ids.shape[1]:], skip_special_tokens=True))These are public Python function benchmarks and cannot guarantee performance on unseen code. Filtering by exact word spans cannot rule out all semantic overlap. Base model: Qwen team, Apache 2.0. Training data: NVIDIA OpenCodeInstruct, CC BY 4.0. The adapter weights were verified byte-for-byte against the training output before upload.
from peft import PeftModelfrom transformers import AutoModelForCausalLM, AutoTokenizer base = AutoModelForCausalLM.from_pretrained( "Qwen/Qwen3-1.7B-Base", revision="ea980cb0a6c2ae4b936e82123acc929f1cec04c1", dtype="auto", device_map="auto",)adapter = "Eternity5551/Qwen3-1.7B-Python-Code-LoRA-SFT"model = PeftModel.from_pretrained(base, adapter).eval()tokenizer = AutoTokenizer.from_pretrained(adapter) prompt = '"""\nWrite a Python function double_even(nums) that doubles only even integers.\n"""\n\n'inputs = tokenizer(prompt, return_tensors="pt").to(model.device)output = model.generate(**inputs, do_sample=False, max_new_tokens=512, pad_token_id=tokenizer.eos_token_id)print(tokenizer.decode(output[0, inputs.input_ids.shape[1]:], skip_special_tokens=True))