PorTAL refit for Inkling

This is a native PorTAL artifact refitted onto thinkingmachines/Inkling. Its 14-task latent table and canonical LoRA-generating core were learned jointly from Qwen3-1.7B and Qwen3-4B and frozen during refitting. Only a fresh Inkling alignment was trained.

Inkling is loaded through its multimodal wrapper for text-only execution. The artifact records 132 exact q/v projection targets across its language-model decoder, including their heterogeneous input and output dimensions, and generates rank-8 LoRA factors only for those paths.

Evaluation

One seed was evaluated on the pinned 14-task validation suite using continuation log-probability divided by character length (acc_norm). Gold continuation token-mean NLL was tracked separately.

The selected checkpoint was epoch 4. The independently minimum-NLL checkpoint was epoch 1 with macro gold NLL 1.8753. Across the same tasks, independently trained rank-16 full-target LoRA baselines averaged 0.8610. Relative to the frozen base, PorTAL retained 104.4% of their aggregate lift.

These are research benchmark results for this exact artifact and evaluation recipe, not a general performance guarantee. The complete epoch history and diagnostics are in metrics.json.

Refit recipe

• Base: thinkingmachines/Inkling at 828496eeae4c243ff1a22f7f28ff83694f2f7bc9
• Frozen source carrier: RampPublic/portal-qwen3-4b@v0.2.0
• Dataset: RampPublic/portallib-tasks at ffc3c0e44f529bf64a5ae62ed5db090952db97ea
• Refit data: deterministic seeded sample of up to 1,000 examples per task from the full pool
• Evaluation data: leading 1,000 validation examples per task, or all rows for shorter tasks
• Optimization: 5 epochs, batch size 4, alignment LR 2e-5, linear decay with 10% warmup, norm-equalized task gradients, character-normalized choice-loss weight 3, seed 0
• Trainable parameters: target-base alignment only; task latents and canonical core remained frozen
• Checkpoint selection: maximum macro validation acc_norm, with lower gold NLL as the tie-breaker
• Architecture: exact q/v targets, rank 8, alpha 16, task latent 256, layer embedding 32, hidden 512, canonical width 1024

The checked-in recipe is examples/configs/refits/inkling.toml.

Usage

Install portallib>=0.2.1, then load the immutable artifact revision:

from portallib import PortalModel

portal = PortalModel.from_pretrained(
    "RampPublic/portal-inkling",
    revision="v0.2.1",
)
portal.export_peft("rte", "./portal-rte-inkling")

The exported directory is an ordinary PEFT LoRA adapter for the exact base revision recorded above. The artifact is Apache-2.0; the benchmark dataset contains components under multiple upstream licenses documented on its dataset card.