README.md

---
license: other
license_name: personaplex-derivative
base_model: nvidia/personaplex-7b-v1
library_name: peft
tags: [full-duplex, speech, instruction-following, lora, moshi, personaplex]
extra_gated_prompt: >-
  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.
extra_gated_fields:
  Name: text
  Affiliation: text
  Intended use: text
---

# oif — online-instruction-following SFT checkpoints

LoRA adapters for **[nvidia/personaplex-7b-v1](https://huggingface.co/nvidia/personaplex-7b-v1)**
that make the model follow **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 ever speaking it.

Project: [MagicLuke/duplex-online-if](https://github.com/MagicLuke/duplex-online-if) ·
recipe of record: `docs/two_stage_training.md` · per-recipe lab notes: `docs/lab/` ·
RL post-training of the same base: [`MagicLuke/oif-instok-rl`](https://huggingface.co/MagicLuke/oif-instok-rl)

## Contents

One subfolder per published checkpoint, `<run_tag>/checkpoint_<step>/`.

| run | step | what it is | downstream eval |
|---|---|---|---|
| `stage1f_final_ins` | 4000 | **Stage 1 only** — Fisher + Seamless directive injection, never saw a goal card. The checkpoint stage 2 was initialised from, and the peak of the stage-1 conditioning probe | **none** |
| `stage2_ins_stage1f` | 337 | **The shipped SFT checkpoint.** 337 steps of conv-synth goal arcs continued from stage 1 via `lora_init_from` | IF-Bench v1.3 + FDB v1.0 |

Read that eval column before choosing. `stage1f_final_ins/4000` is published as the controlled
stage-1 baseline for "what did stage 2 add" — it has never been benchmarked and it **cannot
initiate speech from a cold prompt** (see below). If you want a usable model, use step 337, or
the RL checkpoints.

## `stage2_ins_stage1f/checkpoint_000337`

The SFT checkpoint of record. `lora_init_from` stage 1's step 4000, then 337 steps on conv-synth
goal arcs; `instr_lr 0.0` (delimiter rows frozen), `instr_dropout 0.0`, lr 1e-6, bs 2 × 8 ranks,
240 s windows. Eval loss 2.097 → **1.597** (text 0.715 → 0.365, audio 1.382 → 1.233).

`lora.safetensors` is 2,328,712,528 bytes, sha256
`6ccafd8707471d89620a2767bec1efe866b4527fb60461e451705730ab5ce4d8`, folded at `--scaling 2.0`.

### IF-Bench v1.3 — 200 tasks, judge `Qwen/Qwen3.8-27B`

| examiner | topic 1–5 | completion | either-side | order | seat-flip ↓ | naturalness |
|---|---|---|---|---|---|---|
| stage2-337 + brain | 3.79 | **49.1%** | 58.4% | 9.5% | **4.6%** | 4.05 |
| stage2-337 brain off | 2.19 | 14.6% | 19.6% | 0.0% | **3.6%** | 4.54 |
| *`oif-instok-rl` gdpo-duplex v1200 + brain* | 4.80 | 85.8% | 91.4% | 47.5% | 42.5% | 3.81 |
| *PersonaPlex-7B base* | 2.19 | 7.6% | 10.6% | 0.0% | 21.5% | 4.45 |

**This checkpoint's real strength is role stability, not completion.** At 3.6–4.6% quote-verified
seat flips it is an order of magnitude steadier than either RL checkpoint (40–66%) and better
than the base model (21.5%), and its naturalness is the highest of our checkpoints. It completes
far fewer tasks. Pick accordingly.

### FDB v1.0 — 727 stimuli

| condition | pause TOR syn/candor ↓ | backchannel TOR/freq/JSD | smooth TOR/lat | interrupt TOR/rating/lat |
|---|---|---|---|---|
| brain off, training card + voice | 0.124 / 0.236 | 0.073 / 0.011 / 0.953 | 0.857 / 0.992 | 0.960 / 2.312 / 0.908 |
| brain off, VARF0 + FDB prompts | 0.153 / 0.245 | 0.018 / 0.056 / 0.795 | 0.756 / 0.007 | 0.950 / 4.037 / 0.859 |
| raw adapter, NATF2 + FDB prompts | 0.000 / 0.009 | 0.255 / 0.045 / 0.786 | 0.244 / 0.997 | 0.930 / 3.914 / 0.977 |
| *PersonaPlex (ours, VARF0)* | 0.642 / 0.718 | 0.418 / 0.176 / 0.700 | 0.983 / 0.020 | 0.935 / 4.684 / 0.214 |

**The finetune regressed turn-taking on this benchmark** — smooth-turn TOR fell 0.96 → 0.24 in
the raw-adapter condition. Low pause TOR is good (it holds silence), but taken together with the
smooth-turn collapse the picture is a model that has become reluctant to take the floor.

### Training-data overlap with IF-Bench — measured

**Disclosure:** conv-synth v1 was seeded from the 200 FDB-v2 staged tasks and split *by scenario*
into 180 train / 10 val / 10 test — and those 200 scenarios **are** the 200 IF-Bench task ids. So
stage 2 trained on scenarios matching 180 of the 200 benchmark tasks. (Stage 1 did not; its
corpus is Fisher + Seamless.)

Verified completion on the 20 genuinely held-out scenarios minus the 180 trained ones, bootstrap
CI over tasks:

| arm | held-out (n=19) | trained (n=172) | gap | 95% CI |
|---|---|---|---|---|
| stage2-337 brain off | 18.4% | 14.8% | +3.6 | [−5.7, +14.3] |
| stage2-337 brain on | 50.0% | 49.9% | +0.1 | [−13.0, +13.4] |

**No memorization advantage is detectable** — both CIs span zero and the point estimates are
flat-to-positive. The overlap is real in provenance but does not show up in the score. The test
has limited power at n≈20 (roughly ±10 points), so it rules out a large effect, not a small one.
Held-out ids: `fdb-v2-IF-bench/benchmark/oif_splits/convsynth_overlap.json`; score both halves
with `slurm/arm_compare.py --subsets`.

## `stage1f_final_ins/checkpoint_004000`

Stage 1 of the two-stage recipe, stopped for publication at the step stage 2 branched from.

- **Base** `personaplex-7b-v1-native-dq8-instr32002` — dep_q sliced 16→8, text vocab 32000→32002
  for the two directive markers. The adapter does **not** load onto stock PersonaPlex unchanged;
  see the repo's `scripts/convert/`.
- **Data** `instr_full/v2` — Fisher-train + Seamless improvised + Seamless naturalistic, weighted,
  240 s windows. 221,247 examples / 1,652,922 injections.
- **Recipe** LoRA r512 + `depth_rank` 64, scaling 2.0 · lr 2e-6, `instr_lr` 2e-5 ·
  `instr_dropout` 0.5→0.3 · `instr_near_p` 0.7 / `near_frames` 25 · `text_padding_weight` 0.3 ·
  `first_codebook_weight_multiplier` 100 · bs 8.

### Measured behaviour, including what is wrong with it

Directive probe on 200 held-out Fisher dialogues (1,850 scored injections), `wrong − correct`
cross-entropy, base-corrected:

| checkpoint | base | 1000 | 2000 | 3000 | **4000** | 5000 | 6000 | 7000 |
|---|---|---|---|---|---|---|---|---|
| base-corrected Δ | — | +0.088 | +0.085 | +0.088 | **+0.100** | +0.091 | +0.092 | +0.081 |

**Steering saturates at the first saved checkpoint.** Steps 1000 → 7000 move the conditioning
delta by less than 0.02, and the peak is at 4000, not at the end.

**This checkpoint cannot initiate speech from a cold prompt.** Free-running from a voice+persona
prompt with a real partner talking and no teacher-forced history, it produces **0 words in 120 s**
where the base model produces 41. The same checkpoint looks *better than base* under any fixture
that supplies 30 s of real conversation first (conditionality delta +0.257 vs base's +0.116) — it
can continue speech but not start it. Treat teacher-forced numbers for this checkpoint as
measuring conditioning only, never conversational competence.

It is published because it is the controlled stage-1 baseline for "what did stage 2 add", not
because it is a usable conversational model on its own.

**Why `ins_token`.** Every variant that routes the instruction through a trainable projection
measured **at chance**, including on this exact corpus: `add_sum` reached Δ = +0.0002 ± 0.0003
after 4000 steps — correct, wrong, and *no instruction at all* were identical to three decimals,
while its eval loss fell normally. Only mechanisms carrying **real token ids** (`inline`,
`ins_token`) ever learned. That is the single most useful negative result here.

## Usage

Instructions ride as **real text-token spans spliced in as extra timesteps**:

```
span = [32000, *spm.encode("ask them about their job"), 32001]   # <instr_start> … <instr_end>
```

Fold into NVIDIA's base with the tested recipe from
[`duplex-brain`](https://github.com/MagicLuke/duplex-brain):

```bash
huggingface-cli download MagicLuke/oif-instok-sft \
    --include "stage2_ins_stage1f/checkpoint_000337/*" --local-dir adapter
python -m duplex_brain.ckpt.build_checkpoint \
    --hf-repo nvidia/personaplex-7b-v1 \
    --adapter adapter/stage2_ins_stage1f/checkpoint_000337/lora.safetensors --scaling 2.0 \
    --out ckpt/duplex-brain-stage2.safetensors
```

`--scaling 2.0` is the run's `lora_scaling`, applied directly, **not** divided by rank. Keep the
generated `.depq.json` and `.vocab.json` sidecars beside the weights.

## ⚠️ Two failure modes that load cleanly and are silently wrong

1. **The sidecars are not optional.** Loading the `dep_q=8` weights without `.depq.json` silently
   duplicates agent heads into partner heads; without `.vocab.json`, `<instr_start>` silently
   aliases BOS. Keep both files beside the `.safetensors`.
2. **The injection mode is required and cannot be inferred.** An adapter's tensors cannot
   distinguish `ins_token` from `inline`. Always pass `mode=ins_token`; a wrong mode produces
   plausible-sounding, wrong output rather than an error.

Ids ≥ 32000 are not SentencePiece pieces — `id_to_piece(32000)` raises. Render them from the
`.vocab.json` sidecar, never through the tokenizer.

## Caveats

- **Free-run speech quality was never systematically evaluated** during training. The documented
  failure mode for this recipe family is the model going near-mute (13.2 → 3.9 words/clip) *while
  eval loss improves*. Step 337 has downstream benchmark numbers, which is stronger evidence than
  any offline metric here — but run `audio_health` (≥6 seeds; the failure is ~50% per-rollout
  stochastic) before trusting a checkpoint for generation.
- **No directive probe was run on any stage-2 checkpoint.** Whether stage 2 preserved stage-1's
  steering is unmeasured, and stage 2 drops Fisher replay by design.
- **Both runs were cut by a 3-day wall clock.** Stage 1 stopped at 7000 of a planned 10000 steps,
  so the OneCycle schedule annealed to ~4.5e-7 rather than to 0. Stage 2 used `batch_size 2`
  (vs 4 in the reference recipe).
- Stage-1-only models **never saw a conv-synth goal card**. Prompt them with a caller-seat task
  prompt (they are the user side, not an assistant) and expect loose adherence to goal lists.
- Eval delivers *all* instructions deterministically (no dropout), so eval loss is self-consistent
  across checkpoints but not comparable to training loss.
- Trained on Fisher (LDC) and Seamless Interaction; neither corpus is redistributed here.
- Base weights are NVIDIA's PersonaPlex — their licence governs the base model.

## Provenance

Code: `MagicLuke/duplex-online-if`. Training data derived from Fisher (LDC), Seamless Interaction
and the conv-synth goal-arc corpus; instruction annotations from
`MagicLuke/{fisher,seamless}-turn-instructions-v1`. Benchmark outputs:
[`MagicLuke/ifbench-conversations-v1`](https://huggingface.co/datasets/MagicLuke/ifbench-conversations-v1)
and [`MagicLuke/fdb-v1-outputs-v1`](https://huggingface.co/datasets/MagicLuke/fdb-v1-outputs-v1).