fragment-2 (v3.0.2)

A 36.3M-parameter System-One decision model, built by FrameXlabs on the original fragment build 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: choice, score, noul, multi, rank.

Results (measured by the training notebook, held-out splits)

benchmark	fragment-2	Laya (reported)
AG News (choice)	0.9080	0.950
DBpedia-14 (14-way choice)	0.1540	—
Yelp-5 (score)	0.6040	SST-5 0.372
SST-5 (score, zero-shot)	0.3421	zero-shot base 0.362
IMDb (noul, zero-shot)	0.8345	—
multi F1@0.5 (AG News)	0.7860	n/a (no multi type)
rank Kendall tau (Yelp)	0.5773	n/a (no rank type)
avg ECE (calibrated)	0.1971	0.081 (raw 0.213)
avg accuracy	0.5920	typed-decisions 0.766

Sizes: fragment-2 36.3M from scratch vs Laya 421M/322M

Quickstart

The repo ships a self-contained runtime — f3.py (needs only torch + safetensors):

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={
        "urgency": {"type": "score",
                    "instructions": "How urgent is this request?",
                    "criteria": ["not urgent", "soon", "critical or blocking"]},
        "refund": {"type": "noul",
                   "instructions": "Does the user explicitly request a refund?"},
        "dept": {"type": "choice",
                 "instructions": "Which department should handle this?",
                 "criteria": {"billing": "invoices, payments, refunds",
                               "technical": "bugs, outages, errors",
                               "other": "everything else"}},
        "claims": {"type": "multi",
                   "instructions": "Which of these claims hold for this message?",
                   "criteria": ["the customer is angry",
                                "a refund is requested",
                                "this is about march billing"]},
    })
print(res["answers"]["refund"]["noul"])      # calibrated P(yes)
print(res["answers"]["urgency"]["score"])    # expected level + distribution

Architecture

	fragment-1 (v2.29)	fragment-2 (v3.0.2)
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	36,332,161
question types	3	5 (+ multi, rank)
options per question	5	16 (trained 14-way)

Training: supervised warmup (cross-entropy / BCE / Plackett-Luce over ~470k typed items from SST-2, BoolQ, Amazon polarity, AG News, DBpedia-14, Yelp-5) → RLCD v2 (GRPO-style Gaussian logit-noise exploration, strictly proper scoring rules as reward — log score, RPS, Brier, PL log-likelihood) → per-type temperature calibration on a held-out split.

Honest limits

• English-first. The tokenizer carries a char-level fallback alphabet for other scripts (encodable, but no sub-word merges outside Latin) — a multilingual fragment on this same build is next.
• Zero-shot numbers are measured on IMDb / SST-5, which were never trained on; they are reported, not guaranteed, on arbitrary domains.
• Fine-tuned Laya checkpoints still win some benchmarks (e.g. Banking77 0.870 vs 0.425); fragment-2's case is: from-scratch, ~9x smaller, 5 question types, calibrated out of the box, fully reproducible.
• Scores are ordinal rubric answers, not generation — it will not write prose, by design.

Lineage

fragment-1 v2.29 (8.98M, CPU-trained) → fragment-2 v3.0.2 (same recipe, scaled, GPU-trained). One build, no graveyard: the notebook is the complete, reproducible pipeline.

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