longarmd/CLIFT

CLIFT: Contextual Learning across Inference, Format, and Transfer

A structured benchmark of 5,160 synthetic instances that stress-test whether models learn latent rules from context.

---

At a glance

Instances	5,160
Design	Full factorial over task × format × application × difficulty, with 10 i.i.d. draws per cell
Seed	42
Language	English prompts and instructions
Modality	Text (completion-style prompt → string target)

Load in Python

from datasets import load_dataset

ds = load_dataset("longarmd/CLIFT", split="train")
# Each row: prompt, target, task, format, application, difficulty, latent_structure, ...

If the Hub loader is misconfigured, use the Files tab JSONL and stream with json.loads per line—the schema matches the table below.

---

Why CLIFT?

Modern LMs are often evaluated on fixed formats or single-shot skills. CLIFT targets a narrower but critical capability: contextual learning—inferring a hidden structure from the prompt (examples, traces, or specs), then answering under controlled variation along three axes:

1. Inference — What must be learned (lookup rules, algorithms, spatial transforms, small dynamical systems, …).
2. Format — How that knowledge is presented (demonstrations, natural language, execution traces, formal specs).
3. Transfer / application — How the model must use it (forward prediction, inverse reasoning, articulation, OOD probes, planning, structural probes—task-dependent subsets).

Together, these axes yield a dense evaluation matrix suited for diagnostics, ablations, and comparing training or prompting strategies—not a single leaderboard score in isolation.

---

Task families

Instances are grouped into four families spanning 10 canonical tasks:

Family	Tasks
Functional mappings	lookup_table, arithmetic_rule, conditional_rule
Algorithmic	insertion_sort, max_subarray, binary_search, naive_string_matcher
Spatial	spatial_translation
Dynamic structures	affine_dynamics_2d, register_machine_2d

Formats (all tasks use this set where applicable): demonstration, natural_language, trace, formal_spec.

Difficulty: integer levels 1, 2, 3 (structure complexity scales with level).

Application varies by task (e.g. affine/register tasks use dedicated OOD-suffixed probes). The shipped matrix matches the open-source generator defaults in clift.common.

---

Dataset structure

Data are distributed as JSONL: one JSON object per line, Hugging Face–friendly and line-diffable. A companion manifest.json (in the source repo) records generator kwargs, expected row count, and a SHA-256 over the canonical payload for integrity checks.

Fields (per instance)

Field	Type	Description
instance_id	int	Stable index within the snapshot
task	string	One of the 10 canonical task names
format	string	Presentation format
application	string	Probe / application axis
difficulty	int	1–3
prompt	string	Model input (completion-style)
target	string	Reference answer (exact match is the primary check)
latent_structure	object	Gold structure for analysis & tooling (not shown to the model)
instruct	bool	Whether instruct/chat-style export was used
messages	array (optional)	OpenAI-style chat turns, when enabled at export
metadata	object (optional)	Extra fields when present

Note: latent_structure is intentionally included for research and scoring pipelines. Treat it as held-out supervision for training—do not condition generation on it unless your experimental design explicitly allows it.