fragment-2 (v3.0.0)

A ~36M-parameter System-One decision model by FrameXlabs — built on the original Fragment build (fragment-1 v2.29) and trained entirely from scratch: own BPE tokenizer, own encoder, own RLCD loop. No pretrained weights, no fine-tune of Laya or any other model.

Give it a state (any text) and typed questions; it returns typed answers with calibrated probabilities in a single forward pass — no autoregression, no text generation, no chain-of-thought, nothing to parse and nothing to hallucinate.

Five question types (Laya has 3): choice, score, noul, multi, rank — with up to 16 options per question (trained 14-way on DBpedia).

Weights

This repository is the release target of the training notebook — the first trained checkpoint is published here automatically when the notebook is run (Stage 10, "automatic release"):

fragment-2-colab.ipynb — one free Google Colab T4 run (~2.5 h) trains the model from scratch and publishes weights + tokenizer + config + measured model card to this repo (and to GitHub). f3.py and f3_config.json are already here: the runtime is ready the moment weights land.

vs Laya (Convai Innovations, Sep 2026)

	Laya	fragment-2
Origin	fine-tuned ModernBERT / mmBERT	from scratch
Params	421M (EN) / 322M (multi)	~36M
Question types	3	5 (+ multi, rank)
Options per question	weak beyond 20	up to 16, trained 14-way
Calibration	raw ECE 0.213 → 0.081 tuned	RLCD + per-type temperature
Zero-shot	0.362 reported	measured on IMDb + SST-5 (never trained on)
Reproducibility	weights only	weights + tokenizer + full notebook

Honest note: Laya-multilingual serves 100+ languages; fragment-2 is English-first (char-level fallback alphabet for other scripts in the tokenizer). Multilingual is the next fragment release, on this same build.

Architecture (the original F2Net, scaled and modernized)

	fragment-1 (v2.29)	fragment-2 (v3)
layers / width	6 · d=256	10 · d=384
attention	4 heads, learned pos	6 heads, RoPE
FFN	GELU 1024	SwiGLU 1536
vocab / context	16,384 / 192	32,768 / 256
params	8.98M	~36M

Training = the original three-stage recipe, extended:

1. Supervised warmup — CE / BCE / Plackett-Luce over ~470k typed items (SST-2, BoolQ, Amazon polarity, AG News, DBpedia-14, Yelp-5, plus new multi + rank constructions)
2. RLCD v2 — GRPO-style Gaussian logit-noise exploration with strictly proper scoring rules as reward (log score, RPS, Brier, PL log-likelihood); honest probabilities are the unique reward maximiser
3. Per-type temperature calibration on a held-out split

Quickstart (after weights land)

from f3 import Fragment
m = Fragment.from_pretrained("FrameXlabs/fragment-2")
res = m.decide(
    state="Hi, we were billed twice for March. Please refund the duplicate today.",
    questions={
        "refund": {"type": "noul",
                   "instructions": "Does the user explicitly request a refund?"},
        "urgency": {"type": "score",
                    "instructions": "How urgent is this request?",
                    "criteria": ["not urgent", "soon", "critical or blocking"]},
    })
print(res["answers"]["refund"]["noul"])   # calibrated P(yes)

Lineage

fragment-1 v2.29 (8.98M, CPU-trained) → fragment-2 v3 (same recipe, scaled, GPU-trained). One build, no graveyard.

— FrameXlabs, a group of students who want to build something big.