README.md

metadata

license: apache-2.0
tags:
  - code
  - agent
  - merged
  - fine-tuned

Primus

A coding model that fixes what developers actually complain about.

Primus is a 295B-parameter coding model fine-tuned using the APEX method stack. Unlike frontier models optimized for benchmark scores, Primus is optimized for the five complaints developers have about AI coding assistants:

1. Regression-free repair — fixes the bug without breaking adjacent functionality
2. Scope discipline — changes only what you asked for, not 47 files
3. Ask-vs-guess — asks clarifying questions instead of guessing wrong
4. API honesty — never invents functions or libraries that don't exist
5. Terseness — minimal diffs, no essays

Method

• Training: LoRA SFT → TIDE-DPO → RLVR sprint
• Data: 31,453 decontaminated samples (priors, CAT units, behavior demonstrations, TIDE pairs)
• Hardware: 8× RTX PRO 6000 (764GB VRAM)

The APEX Stack

• CAT (Consequence-Augmented Training): Predicts blast radius before acting
• TIDE (Step-Anchored DPO): Assigns credit/blame at the exact causal step
• SENTINEL: Anti-reward-hacking verifier
• AEGIS: Adaptive best-of-N at inference (safest effective option)
• SONAR: External memory for long-horizon tasks

Usage

from transformers import AutoModelForCausalLM, AutoTokenizer

model = AutoModelForCausalLM.from_pretrained(
    "kovacreations/primus",
    trust_remote_code=True,
    device_map="auto",
    torch_dtype="bfloat16"
)
tok = AutoTokenizer.from_pretrained("kovacreations/primus", trust_remote_code=True)

messages = [
    {"role": "system", "content": "You are Primus, a coding assistant. Write clean, minimal code."},
    {"role": "user", "content": "Fix the null pointer in auth.py"}
]
prompt = tok.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
inputs = tok(prompt, return_tensors="pt").to(model.device)
out = model.generate(**inputs, max_new_tokens=2048)
print(tok.decode(out[0], skip_special_tokens=True))

Scorecard

Pre-registered 15-axis scorecard with all results published (wins and losses):

• 10 public benchmarks (Terminal-Bench, SWE-bench, etc.)
• 5 DevPain axes (regression, scope, ask, API honesty, terseness)

Research Paper

See PRIMUS_PAPER.md for the full method description, training details,

License

Apache 2.0. Decontaminated training data. All losses published.

---

Primus: first, do no harm to the codebase.