Mossez-100M-Coder-Base

Mossez-100M-Coder-Base

Mossez-100M-Coder-Base is an experimental 100M-parameter code completion and fill-in-the-middle model continued-pretrained from mossez-systems/Mossez-100M-Base. It is a base model, not a chat or instruction-following assistant.

Model details

The tokenizer extends the Mossez-100M-Base vocabulary with seven single-token chat/FIM markers. Existing token IDs were not changed. The FIM markers are <|fim_prefix|> (32004), <|fim_middle|> (32005), and <|fim_suffix|> (32006).

Usage

from transformers import AutoModelForCausalLM, AutoTokenizer

model_id = "mossez-systems/Mossez-100M-Coder-Base"
tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(model_id, device_map="auto")

prompt = "def fibonacci(n: int) -> list[int]:
"
inputs = tokenizer(prompt, return_tensors="pt").to(model.device)
output = model.generate(**inputs, do_sample=False, max_new_tokens=96)
print(tokenizer.decode(output[0], skip_special_tokens=True))

For fill-in-the-middle, render the prompt as <|fim_prefix|>{prefix}<|fim_suffix|>{suffix}<|fim_middle|>.

Training and evaluation

The model consumed 39,997,440 tokens in 9,765 finite optimizer steps without corpus wraparound. Packed validation loss decreased monotonically from 2.572834 to 1.488147. See TRAINING_REPORT.md, EVALUATION.md, and DATASET_ATTRIBUTION.md.

The released model.safetensors SHA-256 is aba529bf10ad9f3acb5294c8bc2b4c93d20d25c6cff3a235a8659503b9ac1837.

Limitations

This small research model is not production-ready. It can emit malformed or insecure code, wrong constants, hallucinated APIs, repetition, and early EOS. Its 1,024-token context is short, and the evaluation suite is narrow. Validate, test, and sandbox every output. Do not use generated code without review.