--- license: apache-2.0 base_model: Qwen/Qwen2.5-Coder-0.5B-Instruct language: - en pipeline_tag: text-generation library_name: transformers tags: - code - rust - lindot --- # Lin-1.1 Epsilon **Lin-1.1 Epsilon** is a 0.5B-parameter Rust code model from LindotAI, fine-tuned from [Qwen2.5-Coder-0.5B-Instruct](https://huggingface.co/Qwen/Qwen2.5-Coder-0.5B-Instruct). It is small enough to run on a laptop CPU and is meant for quick completions, simple functions and generate-and-test loops. Its bigger sibling is [Lin-1 Locus](https://huggingface.co/LindotAI/Lin-1-Locus) (4B). 1.1 replaces Lin-1 Epsilon Preview. It improves pass@1 on HumanEval-Rust and MBPP-Rust, our private holdout (+4.1 points) and LBPP-Rust. The Preview stays available at [LindotAI/Lin-1-Epsilon](https://huggingface.co/LindotAI/Lin-1-Epsilon). ![Benchmarks](Lin-1.1-Epsilon-benchmarks.png) ## Benchmarks All numbers were measured by us with the same harness for every model. The benchmarks are Rust ports; generated code is compiled and run against the tests. pass@1 and pass@10 are computed from 10 samples per problem. | Model | HumanEval-Rust p@1 / p@10 | MBPP-Rust p@1 / p@10 | Holdout (Rust) | LBPP-Rust | |---|---|---|---|---| | **Lin-1.1 Epsilon (0.5B)** | **49.4** / 63.3 | 43.9 / **54.4** | **34.4** | **4.73** | | Lin-1 Epsilon Preview (0.5B) | 48.4 / **65.2** | 41.3 / 54.2 | 30.3 | 4.11 | | Temper 2 Ion (0.5B) | 48.1 / 57.6 | **45.5** / 53.8 | 31.0 | 4.50 | | Temper 1 0.5B | 36.3 / 46.3 | 39.3 / 46.9 | — | — | | ThoxEdge-RustCoder-0.5B | 28.7 / 44.8 | 33.2 / 43.1 | — | — | | Qwen2.5-Coder-0.5B-Instruct (base) | 23.8 / — | — | — | — | | Qwen3.5-0.8B | 10.1 / 22.1 | 14.5 / 25.5 | — | — | *Holdout* is a set of 1,726 Rust tasks with tests that is never used for training. *LBPP-Rust* (129 tasks) is a harder problem set than MBPP; small models solve few of its problems, so differences there are small in absolute terms. ## Quick start ```python from transformers import AutoTokenizer, AutoModelForCausalLM model_id = "LindotAI/Lin-1.1-Epsilon" tok = AutoTokenizer.from_pretrained(model_id) model = AutoModelForCausalLM.from_pretrained(model_id, torch_dtype="auto", device_map="auto") messages = [{"role": "user", "content": "Write a Rust function that checks whether a string is a palindrome."}] inputs = tok.apply_chat_template(messages, add_generation_prompt=True, return_tensors="pt").to(model.device) out = model.generate(inputs, max_new_tokens=512, temperature=0.2, do_sample=True) print(tok.decode(out[0][inputs.shape[1]:], skip_special_tokens=True)) ``` GGUF builds (Q8_0, Q4_K_M) are in [LindotAI/Lin-1.1-Epsilon-GGUF](https://huggingface.co/LindotAI/Lin-1.1-Epsilon-GGUF). ```bash llama-cli -hf LindotAI/Lin-1.1-Epsilon-GGUF:Q8_0 -p "Write a Rust function that reverses a Vec in place." ``` Recommended sampling: temperature 0.2 for single answers, 0.8 when sampling several candidates and testing them. ## Training 1. Supervised fine-tuning of Qwen2.5-Coder-0.5B-Instruct on curated, compile-checked Rust. 2. Reinforcement learning (GRPO) with a test-pass reward, then a weight average of the SFT and RL checkpoints (Lin-1 Epsilon Preview). 3. New in 1.1: five further GRPO rounds on Rust tasks with unit tests, each starting from the previous round; the last four rounds used 2,000 new, previously unseen tasks each. The reward uses a pass@k-style advantage. HumanEval-Rust pass@10 is 1.9 points lower than in the Preview; every other score is higher. All training data was decontaminated against the evaluation sets (HumanEval, MBPP, LBPP and the holdout) with a 12-gram overlap filter and exact matching. ## Limitations - A 0.5B model makes more mistakes than larger ones on multi-step logic; LBPP-level problems are mostly out of reach. - Focused on Rust. Always compile and test generated code. ## License Apache 2.0, same as the base model.