# pip install moshi
# Run the interactive web server
python -m moshi.server --hf-repo "MagicLuke/oif-instok-rl"
# Then open https://localhost:8998 in your browser# pip install moshi
import torch
from moshi.models import loaders
# Load checkpoint info from HuggingFace
checkpoint = loaders.CheckpointInfo.from_hf_repo("MagicLuke/oif-instok-rl")
# Load the Mimi audio codec
mimi = checkpoint.get_mimi(device="cuda")
mimi.set_num_codebooks(8)
# Encode audio (24kHz, mono)
wav = torch.randn(1, 1, 24000 * 10) # [batch, channels, samples]
with torch.no_grad():
codes = mimi.encode(wav.cuda())
decoded = mimi.decode(codes)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 GDPO post-training of PersonaPlex. gdpo-v2f-r4 was trained on rollouts seeded from Fisher (LDC); gdpo-duplex was not. Neither corpus is redistributed here. By requesting access you agree not to attempt to reconstruct or redistribute the underlying corpora.
Log in or Sign Up to review the conditions and access this model content.
oif β online-instruction-following RL checkpoints
GDPO post-training of nvidia/personaplex-7b-v1
for instructions injected mid-conversation by a separate text LLM. A directive is spliced
into the model's inner-monologue stream as extra timesteps (ins_token mode) with the audio
held silent, so the model reads the directive and acts on its next turn without speaking it.
RL optimises the free-run regime directly β it trains on what the model actually produces β which is why these checkpoints, unlike the SFT line, are usable as conversational models.
Project: MagicLuke/duplex-online-if Β·
lab notes: docs/lab/rl-R1-gdpo-duplex.md, docs/lab/rl-R7-v2f-r4.md Β·
SFT line: MagicLuke/oif-instok-sft
Contents
One subfolder per published checkpoint, <run_tag>/v<step>/.
| run | step | pool | reward | what it is |
|---|---|---|---|---|
gdpo-duplex |
1200 | conv-synth goal arcs | v1 (4-term) | Highest task-completion checkpoint we have. Beats gpt-realtime-2.1 on IF-Bench completion |
gdpo-v2f-r4 |
600 | Fisher-only, goalless persona | v2 (8-dim, VAD) | Best turn-taking checkpoint we have on FDB v1.0 |
Of 60 launched recipes in this project, these two and one SFT checkpoint are the only three with downstream benchmark numbers. Everything else was scored on in-house metrics only, and is not published.
These are not a curve. Both were initialised from the base native checkpoint, never from
an SFT adapter, and they are alternative recipes rather than successive steps. Do not stack
them, and do not apply either on top of oif-instok-sft.
Results
IF-Bench v1.3 β 200 tasks, judge Qwen/Qwen3.8-27B, examinee fdb-assistant-unaware
"brain on" = steered live by a second LLM; "brain off" = the same weights unsteered. Completion is verified stage completion (claimed% in parentheses); seat-flip and order are v1.1 metrics.
| examiner | topic 1β5 | completion | either-side | order | seat-flip β | naturalness |
|---|---|---|---|---|---|---|
gdpo-duplex v1200 + brain |
4.80 | 85.8% (86.4) | 91.4% | 47.5% | 42.5% | 3.81 |
gdpo-duplex v1200 brain off |
4.59 | 61.8% (62.3) | 70.4% | 9.5% | 40.5% | 3.83 |
gdpo-v2f-r4 v600 + brain |
4.33 | 57.9% (58.8) | 69.0% | 10.0% | 66.5% | 3.68 |
gdpo-v2f-r4 v600 brain off |
4.02 | 36.8% (37.4) | 48.9% | 0.0% | 59.5% | 3.96 |
| gpt-realtime-2.1 (reference) | 4.72 | 73.0% | 80.6% | 32.5% | 47.5% | 3.61 |
oif-instok-sft stage2-337 + brain |
3.79 | 49.1% | 58.4% | 9.5% | 4.6% | 4.05 |
| PersonaPlex-7B base | 2.19 | 7.6% | 10.6% | 0.0% | 21.5% | 4.45 |
v600's either-side numbers are the corrected ones (2026-08-25). A dead judge shard had
scored 18 of 50 Safety dialogues as [null, null, null], and the scorer counts an unparsed
verdict as "no" rather than excluding it, while report.json reported missing: []. That cost
the split 25 points and produced a published either-side figure below the strict one β which
is impossible, since either-side is a union. If you rerun the judge, check for null votes.
FDB v1.0 β 727 stimuli, seed 42424242
The two rows ran under different conditions and are not comparable to each other: v1200 under its training card + voice, v600 under VARF0 + the benchmark's own prompts.
| checkpoint | pause TOR syn/candor β | backchannel TOR/freq/JSD | smooth TOR/lat | interrupt TOR/rating/lat |
|---|---|---|---|---|
gdpo-v2f-r4 v600 (VARF0, FDB prompts) |
0.365 / 0.417 | 0.327 / 0.118 / 0.724 | 1.000 / 0.203 | 1.000 / 4.725 / 0.268 |
gdpo-duplex v1200 (training card + voice) |
0.971 / 0.991 | 0.691 / 0.062 / 0.830 | 1.000 / 0.000 | 1.000 / 2.180 / 0.299 |
| PersonaPlex (ours, VARF0) | 0.642 / 0.718 | 0.418 / 0.176 / 0.700 | 0.983 / 0.020 | 0.935 / 4.684 / 0.214 |
Interruption ratings come from a local gemma-4-E4B judge served as gpt-4-turbo, comparable
only across our own rows.
v1200's pause TOR of 0.971/0.991 is the worst row in that table β it takes the floor during a pause almost every time. That is the cost of the reward it was trained on: reward v1 has no term paying the model to stop talking. If you need turn-taking discipline, v600 is the checkpoint; if you need task completion, v1200 is.
Training-data overlap with the benchmark β measured, and it does not move the score
Disclosure: conv-synth v1 was seeded from the 200 FDB-v2 staged tasks and split by
scenario into 180 train / 10 val / 10 test. Those 200 scenarios are the 200 IF-Bench task
ids. So gdpo-duplex (conv-synth pool) trained on scenarios matching 180 of the 200 benchmark
tasks. gdpo-v2f-r4 did not β its pool is Fisher-only β which makes it the control.
We measured the gap rather than assuming it. Verified completion on the 20 genuinely held-out scenarios minus the 180 trained ones, within each arm, bootstrap CI over tasks (n=20 vs 180):
| checkpoint | arm | held-out | trained | gap | 95% CI |
|---|---|---|---|---|---|
gdpo-duplex v1200 |
brain off | 66.2% | 61.3% | +5.0 | [β7.2, +17.1] |
gdpo-duplex v1200 |
brain on | 88.8% | 85.4% | +3.3 | [β3.9, +9.9] |
gdpo-v2f-r4 v600 (control) |
brain off | 36.2% | 36.8% | β0.6 | [β11.8, +10.7] |
gdpo-v2f-r4 v600 (control) |
brain on | 47.5% | 59.0% | β11.5 | [β25.3, +2.2] |
| stage2-337 (SFT, conv-synth) | brain on | 50.0% | 49.9% | +0.1 | [β13.0, +13.4] |
No memorization advantage is detectable. Every CI spans zero, and the direction is the opposite of contamination: the two conv-synth-trained models score higher on scenarios they never saw, while the Fisher-only control scores lower. The honest reading is that the overlap is real in provenance but does not show up in the score.
This test has limited power β 20 held-out scenarios gives roughly Β±10 points, so it rules
out a large memorization effect, not a small one. The held-out ids are enumerated in
fdb-v2-IF-bench/benchmark/oif_splits/convsynth_overlap.json; score them separately with
slurm/arm_compare.py --subsets and report both halves.
Usage
Fold the adapter into NVIDIA's base with the tested recipe from
duplex-brain:
huggingface-cli download MagicLuke/oif-instok-rl --include "gdpo-duplex/v1200/*" --local-dir adapter
python -m duplex_brain.ckpt.build_checkpoint \
--hf-repo nvidia/personaplex-7b-v1 \
--adapter adapter/gdpo-duplex/v1200/lora.safetensors --scaling 2.0 \
--out ckpt/duplex-brain-gdpo1200.safetensors
--scaling 2.0 is the run's lora_scaling, applied directly, not divided by rank. The
build writes <stem>.depq.json and <stem>.vocab.json beside the weights β keep all three
files together. The loader reads the architecture from those sidecars, and without them it
silently expands 8 β 16 depth heads or aliases <instr_start> onto BOS: both load, run, and are
wrong. Ids β₯ 32000 are not SentencePiece pieces β never round-trip them through the tokenizer.
Checksums (the fold logs recorded these before the benchmark ran):
| file | bytes | sha256 |
|---|---|---|
gdpo-duplex/v1200/lora.safetensors |
609,897,768 | d71f286efc142150bfdac17ea15567f7e3ee68620aa8171bb544107be4bb2e7c |
gdpo-v2f-r4/v600/lora.safetensors |
609,897,768 | 1f1581a7f69a5e989b2886ee89f18270811d463a3bc41b78037e4c31c80305d2 |
Recipe
Model β PersonaPlex-7B native, dep_q sliced 16 β 8, text vocab extended 32000 β 32002 for
<instr_start> (32000) / <instr_end> (32001), context 3000 (240 s @ 12.5 Hz). The adapters
do not load onto stock PersonaPlex unchanged.
LoRA β rank 128 on the semantic path, rank 32 on the depth decoder, scaling 2.0,
embeddings frozen. 674 tensors.
Trainer β GDPO (per-dimension group z-scored advantages, arXiv 2601.05242), k3 KL against an
adapter-off reference. 24 GPUs: 3 trainer ranks + 16 rollout workers + 1 ASR/VAD score worker +
judge shards. G=8 samples per group, 6 groups per step, lr 1e-5. Scoring is
nvidia/parakeet-tdt-0.6b-v2 ASR + silero VAD + a Qwen/Qwen3.8-27B judge.
gdpo-duplex v1200 (R1) |
gdpo-v2f-r4 v600 (R7) |
|
|---|---|---|
| pool | conv-synth goal-arc dialogues | Fisher-only, 9,417 train / 300 val, goalless persona |
| reward | v1: C_goal β 0.5Β·role_fail β 0.5Β·meta_talk β 0.25Β·jarring, silent = β1 |
v2, 8 dims: C_goal, fits, seat_ok, meta_talk + VAD overlap / false_takeover / latency / backchannel |
| KL | fixed Ξ² 0.01 | adaptive Ξ², --beta 0.05 --kl-target 0.3, clamp 5.0, floor Ξ²/2 |
| steps | ran to 1,479; v1200 published | stopped at 1,181 by hand; v600 best |
| in-domain | compliance@0.5 0.455 β 0.779 by v600, flat after; jarring 0.375 β 0.211 | compliance 0.579, fits 0.906, latency 1.34 s; no-instruction control flat 0.277 β 0.278, so the true conditioning delta is +0.288 |
Limitations β what is wrong with these checkpoints
- Both flip seats far more than the SFT model. 40β42% (v1200) and 59β66% (v600) of dialogues
contain a quote-verified seat flip, against 4.6% for
stage2-337. RL bought task completion and spent role stability to get it. If persona adherence is what you need, neither of these is the right checkpoint. - v600's reward under-priced the failure it was meant to price. A strict flip-only judge
finds 14 / 22 / 12% flipped turns at v600 / v900 / v1050 versus 4β8% by the training
seat_okβ recall 0.24β0.32, sorole_failat 0.5 was pricing about one third of real seat flips. - v600 drifts monotonically even while KL stays inside its band. Per-3-step motion is 2.5Γ the per-150-step motion, and KL-to-base is 0.23 at v900 versus 0.22 at v1179 β both inside the 0.3 target. A base-only anchor admits the best and worst checkpoints alike, so there is no restoring force toward v600. Later steps of that run are worse, not better.
- Winning in-domain is not winning the benchmark. v600 beats v1200 on the Fisher held-out conditioning delta and loses to it on IF-Bench by 28 points of completion. They were scored on different distributions β goalless Fisher windows versus scripted caller tasks β and neither is simply "better".
- The examiner sometimes adopts the wrong identity. Free-running from a card that names a business, the model has opened with "Good morning, this is Morning Delight Cafe." β i.e. it became the callee named in its own card. Cards should state the speaking role and who opens, in the imperative.
- No as-launched command line survives for R7. The reconstruction is
slurm/runs/rl/env/R7-gdpo-v2f-r4.envand is marked as reconstructed, not as-launched. R1's topology is committed accurately (rl/submit_rl.sh).
Provenance
Code: MagicLuke/duplex-online-if, trainer
rl/. Benchmark harnesses: fdb-v2-IF-bench (IF-Bench) and fdb-v1-repro (FDB v1.0).
Generated dialogues and per-dialogue judgments for the IF-Bench rows above are published as
MagicLuke/ifbench-conversations-v1;
FDB v1.0 outputs as
MagicLuke/fdb-v1-outputs-v1.
gdpo-duplex has no Fisher lineage β conv-synth pool, base init. gdpo-v2f-r4 was trained
on rollouts seeded from Fisher (LDC) and is gated on that basis; both are gated for consistency.
Base weights are NVIDIA's PersonaPlex and their licence governs the base model.
- Downloads last month
- -
# Gated model: Login with a HF token with gated access permission hf auth login