GLM-5.3 is live. Run Z.ai's latest model on Friendli Model APIs. Try it today ➜
gianniiii
Available on FriendliAI
Run this model inference on single tenant GPU with unmatched speed and reliability at scale.
Model Details
Model Provider
gianniiii
Model Tree
Qwen/Qwen2.5-Coder-7B-Instruct
Input Modalities
Output Modalities
Supported Functionality
Held-out set: the 228 problems of LeetCodeDataset's test split, released after mid-2024. Each task gets 10 samples at temperature 0.8 plus one greedy response, graded by the same sandboxed verifier used in training.
| Base model | This model | |
|---|---|---|
| LeetCode held-out pass@1, sampled | 14.1% | 21.9% |
| LeetCode held-out pass@1, greedy | 14.9% | 20.2% |
| LeetCode held-out pass@5 | 24.3% | 28.2% |
| LeetCode held-out pass@10 | 28.5% | 28.9% |
| pass@1 Easy / Medium / Hard (48 / 101 / 79 tasks) | 44.6 / 7.9 / 3.4% | 56.0 / 17.1 / 7.3% |
| HumanEval+ pass@1 / pass@10 | 83.0 / 93.9% | 85.6 / 90.2% |
| MBPP+ pass@1 / pass@10 | 68.1 / 85.2% | 72.5 / 82.8% |
| Algorithm | GRPO: 8 responses per prompt, advantage = (reward − group mean) / group std, PPO-clip 0.2, no KL term, no critic |
| Reward | 1 if every unit test passes, else 0. The code runs sandboxed (user, network, PID and mount namespaces; 6 CPU-seconds per response) |
| Data | LeetCodeDataset train split minus 611 tasks every earlier response had solved: 2,030 tasks. 90 steps × 64 prompts ≈ 2.8 passes |
| Hyperparameters | lr 1e-6 (constant), AdamW, 64 prompts per step, 4 optimizer steps per batch, grad clip 1.0, temperature 1.0, max response 1,024 tokens |
| Stack | veRL 0.9.1 (sync trainer), vLLM 0.24 rollouts, FSDP2, 4 × GH200. vLLM's log-probs were reused as the old policy's |
| Precision | Trained with fp32 master weights and bf16 compute; exported in bf16 |
Use the base model's chat template. Training prompts gave the problem statement and LeetCode's starter code, and ended with: "Write your solution in Python 3 and put the complete code in a single ```python block."
python
from transformers import AutoModelForCausalLM, AutoTokenizer repo = "gianniiii/Qwen2.5-Coder-7B-Instruct-GRPO-LeetCode"tok = AutoTokenizer.from_pretrained(repo)model = AutoModelForCausalLM.from_pretrained(repo, dtype="bfloat16", device_map="auto")messages = [{"role": "user", "content": "Write a Python function two_sum(nums, target) that returns the indices of the two numbers adding up to target."}]inputs = tok.apply_chat_template(messages, add_generation_prompt=True, return_tensors="pt").to(model.device)print(tok.decode(model.generate(inputs, max_new_tokens=512)[0][inputs.shape[1]:], skip_special_tokens=True))from transformers import AutoModelForCausalLM, AutoTokenizer repo = "gianniiii/Qwen2.5-Coder-7B-Instruct-GRPO-LeetCode"tok = AutoTokenizer.from_pretrained(repo)model = AutoModelForCausalLM.from_pretrained(repo, dtype="bfloat16", device_map="auto")messages = [{"role": "user", "content": "Write a Python function two_sum(nums, target) that returns the indices of the two numbers adding up to target."}]inputs = tok.apply_chat_template(messages, add_generation_prompt=True, return_tensors="pt").to(model.device)print(tok.decode(model.generate(inputs, max_new_tokens=512)[0][inputs.shape[1]:], skip_special_tokens=True))