Lin-1-Locus / README.md
kotl1nML's picture
Release weights, GGUF and model card
ae0a605 verified
|
Raw History Blame Contribute Delete
3.6 kB
---
license: apache-2.0
base_model: Qwen/Qwen3.5-4B
language:
- en
pipeline_tag: text-generation
library_name: transformers
tags:
- code
- rust
- lindot
---
# Lin-1 Locus
**Lin-1 Locus** is a 4B-parameter code model from LindotAI, specialized for **Rust**. It is fine-tuned from [Qwen3.5-4B](https://huggingface.co/Qwen/Qwen3.5-4B) on curated, test-verified Rust data and answers directly (no long reasoning traces), which keeps it fast enough for local use.
## Highlights
- Writes idiomatic, compiling Rust: functions, traits, generics, error handling, iterators, unit tests.
- Trained only on code that passed compilation and/or its own tests.
- Strong sampling diversity: high pass@10 makes it a good fit for generate-and-test loops.
- Runs locally: a Q8_0 GGUF is included (~4.5 GB).
## Benchmarks
![Lin-1 Locus vs open code models](benchmarks_wide.png)
![Benchmarks](benchmarks.png)
All numbers measured by us with the same harness for every model (Rust ports of the benchmarks; generated code is compiled and run against the tests).
| Model | HumanEval-Rust p@1 / p@10 | MBPP-Rust p@1 / p@10 | Holdout (Rust) | LBPP-Rust |
|---|---|---|---|---|
| **Lin-1 Locus (4B)** | **73.1 / 86.1** | **63.4 / 73.1** | **58.7** | **27.5** |
| Lin-1 Epsilon Preview (0.5B) | 48.4 / 65.2 | 41.3 / 54.2 | 30.3 | 4.1 |
| Temper 2 Nodel | 73.3 / 81.8 | 64.7 / 70.6 | 55.4 | 26.4 |
| Qwen2.5-Coder-7B-Instruct | 72.4 / 82.6 | 60.5 / 70.0 | β€” | β€” |
| Qwen3.5-4B (base) | 66.2 / 81.4 | 57.7 / 73.4 | β€” | β€” |
| Qwen3-4B | 59.8 / 64.8 | 55.6 / 60.6 | β€” | β€” |
| Qwen2.5-Coder-3B-Instruct | 56.5 / 71.9 | 44.8 / 60.0 | β€” | β€” |
| Qwen2.5-Coder-1.5B-Instruct | 39.0 / 56.1 | 38.3 / 52.1 | β€” | β€” |
| Qwen3.5-2B | 31.7 / 51.3 | 36.0 / 53.2 | β€” | β€” |
*Holdout* is our private set of Rust tasks with tests that was never used for training.
## Quick start
### Transformers
```python
from transformers import AutoTokenizer, AutoModelForCausalLM
model_id = "LindotAI/Lin-1-Locus"
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 returns the n-th Fibonacci number using u128, with tests."}]
inputs = tok.apply_chat_template(messages, add_generation_prompt=True, enable_thinking=False, return_tensors="pt").to(model.device)
out = model.generate(inputs, max_new_tokens=1024, temperature=0.2, do_sample=True)
print(tok.decode(out[0][inputs.shape[1]:], skip_special_tokens=True))
```
### vLLM
```bash
vllm serve LindotAI/Lin-1-Locus --max-model-len 8192
```
### llama.cpp / Ollama (GGUF)
```bash
llama-cli -m Lin-1-Locus-Q8_0.gguf -p "Write a Rust function that parses a CSV line into Vec<String>."
```
Recommended sampling: `temperature 0.2` for single answers, `0.8` when sampling several candidates and testing them.
## Training
1. Supervised fine-tuning of Qwen3.5-4B on Rust instruction data and the model's own test-verified answers (rejection sampling).
2. Supervised fine-tuning on curated Rust from permissively licensed sources: files from *The Stack Edu* (Rust subset) that compile and pass their own unit tests, and compile-checked samples from *Strandset-Rust*.
All training data was decontaminated against the evaluation sets (12-gram overlap filter).
## Limitations
- Focused on Rust; other languages work but are not the target.
- 4B parameters: complex multi-file projects and long-horizon agentic tasks remain hard.
- Always compile and test generated code.
## License
Apache 2.0, same as the base model.