Mossez-100M-Coder-Instruct
Mossez-100M-Coder-Instruct is an experimental 100M-parameter coding instruction model with this weight lineage:
Mossez-100M-Base -> Mossez-100M-Coder-Base -> Mossez-100M-Coder-Instruct.
The general Mossez-100M-Instruct
was used only as a tokenizer, chat-template, release, and inference reference;
its weights were not used as source weights for this model.
Model details
| Property | Value |
|---|---|
| Parameters | 100,098,048 |
| Architecture | Llama-compatible decoder-only Transformer |
| Layers / hidden size | 12 / 768 |
| Query / KV heads | 12 / 4 |
| Context length | 1,024 tokens |
| Vocabulary | 32,007 |
| Objective | Assistant-only SFT loss |
| Weight format | Safetensors, FP32 |
| License | Apache-2.0 |
Usage
from transformers import AutoModelForCausalLM, AutoTokenizer
model_id = "mossez-systems/Mossez-100M-Coder-Instruct"
tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(model_id, device_map="auto")
messages = [{"role": "user", "content": "Write a short Python function that adds two integers."}]
prompt = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
inputs = tokenizer(prompt, return_tensors="pt").to(model.device)
output = model.generate(**inputs, do_sample=False, max_new_tokens=96)
new_tokens = output[0, inputs.input_ids.shape[1]:]
print(tokenizer.decode(new_tokens, skip_special_tokens=True))
Training and evaluation
The model was fine-tuned for one bounded epoch: 660 optimizer steps over 2,640 project-authored examples, using assistant-only loss. Immutable validation and test sets contain 330 examples each across 11 balanced task types. See TRAINING_REPORT.md, EVALUATION.md, and DATASET_ATTRIBUTION.md.
The released model.safetensors SHA-256 is
0aade7d070122633abccd70cc5500e5bc36c7f69fafeadd9f5ee0b5a3e0766bf.
Limitations
This is a small research model, not a reliable or safe production coding assistant. The authored SFT corpus is balanced but narrow and template-heavy, so held-out loss may overstate general-world capability. Expect repetition, incorrect constants, malformed code, hallucinated APIs, weak instruction following, and early EOS. Validate, test, and sandbox every output.