Nagi-Big
System One typed-decision model (choice / score / noul) — Qwen3.5-4B + PiSSA LoRA + Choice-B option scoring. Soft-distilled from teacher ensemble (DeepSeek flash + GPT-5.6-Luna). No text generation; calibrated option probabilities.
Also known as: Track-G / hail-g1 / nagi-t4-qwen-pissa.
Load
from peft import PeftModel
from transformers import AutoModelForCausalLM, AutoTokenizer
tok = AutoTokenizer.from_pretrained("Qwen/Qwen3.5-4B")
base = AutoModelForCausalLM.from_pretrained("Qwen/Qwen3.5-4B", dtype="bfloat16")
model = PeftModel.from_pretrained(base, "nagisanzeninz/nagi-big-v0")
# optional: model = model.merge_and_unload()
Inference recipe: nagi.render.render_nagi_prompt (definition + 2 demos + state + options) → last-token logits over letter ids A,B,C… → softmax over K options.
See scripts/t4_qwen_pissa.py and scripts/t4_leave_task_out.py in nagisanzenin/nagi.
gold_ltout LTO (2700, 6 public unseen tasks)
| Task | n | hard_acc |
|---|---|---|
| Amazon | 500 | 0.948 |
| DBpedia | 500 | 0.900 |
| CoLA | 500 | 0.724 |
| MRPC | 500 | 0.704 |
| 20NG | 400 | 0.630 |
| QQP | 300 | 0.350 |
| macro | 2700 | 0.709 |
3-way (same gold_ltout)
| Nagi-Big | Laya | Jev | |
|---|---|---|---|
| LTO hard | 0.709 | ~0.35 | 0.773 |
| p50 latency | 84 ms (L4 eval) | 33 ms | 501 ms (API) |
| marginal cost | self-host | self-host | API |
Beats Laya by ~+36 pts on LTO. Within 6.4 pts of Jev. Gates: USEFUL (≥0.60) ✅ · NEAR (≥0.70) ✅ · WIN (≥0.75) ❌.
Train
- 25,000 items from
train_t4_final(1,443 schemas: public multi-task + hail-C novel schemas + burst/synth) - PiSSA r=32 α=64
pissa_niter_4dropout=0, lr 1e-4, bs 2, 1 epoch, max-len 768 - Loss: soft-label KL
L = −Σ q log p(no hard warmup) - H100 BF16 · wall 4816s
Sibling
- Nagi-Smol — 421M M2′ dual encoder, ~11 ms, LTO 0.462
- nagi-t4-m2p-v0 — prior T4 M2′, LTO 0.481