treble-400m

A fine-tuned Von checkpoint (ModernBERT-large encoder + Option-Marker head, ~395M parameters) for routing in Treble. It selects among described options and judges conditions in one encoder pass — it does not generate text.

This is a fine-tune/distillation of Von, not a new architecture. Student base: von-1.2.

Resources

• Training, data pipeline, and reproduction: OscarVGit/treble-distill (see docs/training.md and docs/distillation.md)
• Harness this model serves: OscarVGit/treble
• Evaluation write-up: Treble EVALUATION.md (§11)
• Base model: wfzyx/von

Training (distillation)

• Teacher: DeepSeek (deepseek-chat) via the OpenAI-compatible API.
• Student: Von von-1.2.
• Task: Treble's routing primitives — run gate (Noul), run target (Choice), control verb (Choice), control target (Choice), showdown/collision (Noul).
• Data: synthesized registries and phrasings, labelled by the teacher; mixed with ~20k replay rows from the base corpus.
• Recipe: 1 epoch, lr 1.5e-5, batch 4 × grad-accum 4, --independent_options --digit_split; confidence refit with von calibrate.

Disclosure: labels were produced by a closed teacher model (DeepSeek); this checkpoint is a derivative of its outputs and is released with that disclosure. Do not redistribute the labelled training JSONL.

Evaluation

Measured on Treble's 69-case routing suite (development/regression eval), 3 runs, stable:

model	correct	wrong	escalated
base Von 1.2	61/69	7	1
DeepSeek teacher	67/69	2	0
treble-400m	67/69	1	1

safe=true, no unrequested executions. ~90 ms/decision locally vs ~1.7 s for the hosted teacher. Per-case agreement with the teacher: 94%.

Overlap disclosure (important). This is a development suite, not a sealed holdout. 10 of its 69 cases (13 probes) share an exact (state, question) with the training/val data, and its phrasing informed data iteration. Excluding those:

model	uncontaminated subset
treble-400m	57/59
DeepSeek teacher	57/59

Teacher parity holds on the uncontaminated subset, but the 67/69 headline should not be read as generalization to unseen phrasings. A sealed evaluation set is pending; treat both numbers as directional on n≈60–70.

Usage

pip install von-sdk
pipx install git+https://github.com/OscarVGit/treble   # the harness

# serve this checkpoint
VON_CHECKPOINT_DIR=/path/to/treble-400m von serve

# point Treble at it
export TREBLE_DECISION_URL=http://127.0.0.1:8765/v1/systemone
treble "show me the build procedure"

Files

File	Purpose
model.safetensors	ModernBERT-large encoder (required to serve)
option_marker.pt	Option-Marker scoring head
marker_calibration.json	calibration map + independent_options (do not drop)
config.json, tokenizer.json, tokenizer_config.json	HF config/tokenizer

All six are required for calibration and serving. option_marker.pt alone is not enough: the backend loads the encoder via AutoModel.from_pretrained, which needs model.safetensors.

Limitations

• English only; selects/scores, never generates.
• Scoped to Treble's routing primitives, not general classification.
• Evaluated on a small internal suite; see the overlap disclosure above.
• Keep --independent_options / --digit_split consistent with the training run.

License

Apache-2.0, following Von and ModernBERT. Teacher-derived labels are disclosed; verify the teacher API's terms before redistributing derived weights.