Blaster-Think
Blaster-Think is a snapshot of the v35 hearing ship tip (checkpoint-116000) plus the Loop 2 typed-think decode contract. Weights are the same merged serving export as live Blaster at that tip. Thinking is not extra LoRA in this dump: the gateway prefills Answer: <think> on cued math and strips the span before TTS.
This repo is a separate Hub model so later GRPO / dual-LoRA work cannot overwrite the hearing ship:
| Do not overwrite | This snapshot |
|---|---|
molmo-audio-serving-diar-d | this repo |
molmo-audio-lora-diar-d | molmo-audio-lora-blaster-think |
SERVE_TIP on the live box stays outputs/molmo-audio-lora-diar-d-v35/checkpoint-116000. Think CE at 117500 was aborted (hearing dual-eval HEARING_NO_SHIP). Do not serve 117500.
What’s inside
Same files as the diar-d serving merge (model-*-of-*.safetensors, audio_modules.pt, tokenizer, remote-code sources), plus:
| Path | Purpose |
|---|---|
serving/think_scaffold.py | Loop 2 P0/P1 cue, 192-token body, force-close, loop XOR |
eval/think_eval_116000_gsm_192.json | Scaffolded GSM heldout at this tip |
SNAPSHOT.json | Pin: tip id, parent, abort notes |
Think eval (this tip, scaffolded GSM, n=20)
| Gate | Result |
|---|---|
Format (<think>…</think>) | 20/20 |
| Exact | 8/20 (worded finals and a few real misses; not a ship gate for think) |
| Unearned correct | 0 |
| Uncued leak | 0 |
Exact is a format contract + weak arithmetic, not a reason to invert training from base Molmo-7B-D.
Intended use
- Restore this exact 116000 merge if GRPO or a second think LoRA goes wrong
- Research on cued typed thinking without replacing live Blaster
License
CC BY-NC-SA 4.0. Private unless the owner changes visibility.
Marketing name: Blaster-Think. Technical Hub id: molmo-audio-serving-blaster-think.