You need to agree to share your contact information to access this model

This repository is publicly accessible, but you have to accept the conditions to access its files and content.

These adapters are trained on Fisher (LDC) and Seamless Interaction. The corpora themselves are not redistributed here. By requesting access you agree not to attempt to reconstruct or redistribute the underlying corpora.

PersonaPlex OIF `ins_token` — retired, kept as a citation target

This was the first release of the online-instruction-following adapters. It is superseded. It stays published, under this exact id, because two already-published datasets record it as their examiner_model_id — retiring the id would break the provenance chain of work that has already shipped.

Go here instead:


SFT checkpoints	`MagicLuke/oif-instok-sft` — stage 1 and stage 2, with benchmark numbers and the training-overlap measurement
RL checkpoints	`MagicLuke/oif-instok-rl` — GDPO post-training; higher task completion than anything here

What is still in this repo

stage2/checkpoint_000337/ — the adapter the published benchmark runs used, kept so the citation resolves to real weights. It is byte-identical to oif-instok-sft/stage2_ins_stage1f/checkpoint_000337/ (sha256 6ccafd87…5ce4d8). Prefer the copy in oif-instok-sft, which carries the current model card.
probe/ — directive-probe results for the stage-1 curve, the evidence behind the claim that steering saturates by step 1000.
configs/ — the exact args.yaml of each run, and the LM config.
base/*.depq.json, base/*.vocab.json — sidecars for the vocab-extended base. The base weights were never shipped here; rebuild them (dep_q 16 → 8, text vocab 32000 → 32002) as documented in oif-instok-sft.

What was removed, and how to get it back

The 13 ablation checkpoints (stage 1 steps 1000–7000, stage 2 steps 50–300; ~28 GB) were removed in the retirement commit. They are not gone — the blobs remain in this repo's git history and can be restored from the commit before it. stage1f/checkpoint_004000 also lives on as oif-instok-sft/stage1f_final_ins/checkpoint_004000.

None of them was ever benchmarked downstream, and the probe results they support are kept in probe/, which is why they were the ones dropped.

Correcting the previous card

The card this replaces said "Private / restricted … Do not redistribute these weights or make this repo public without an LDC licensing decision" while the repo was, in fact, public. That was a contradiction, not a policy.

The actual position: these are LoRA deltas, not corpus. Fisher and Seamless Interaction are never redistributed here, and access is manually gated so that it is recorded. That posture is unchanged in the successor repos. NVIDIA's licence governs the base model.

Provenance

Code: MagicLuke/duplex-online-if. Datasets citing this id: MagicLuke/ifbench-conversations-v1, MagicLuke/fdb-v1-outputs-v1.

Downloads last month: -

Inference Providers NEW

This model isn't deployed by any Inference Provider. 🙋 Ask for provider support

Model tree for MagicLuke/personaplex-oif-ins-token-v1

Base model

kyutai/moshiko-pytorch-bf16

Finetuned

nvidia/personaplex-7b-v1

Adapter

(14)

this model