Tiny Interjection Model Alpha
Tiny Interjection Model Alpha (TIM) is an experimental floor-control adapter for typed chat. Given a timestamped event stream, it chooses wait, respond, interject, or continue and may emit a short assistant message.
This repository preserves the reviewed adapter from the earlier Nebius L40S virtual-machine experiment. It is not the ephemeral adapter produced by the later credentialless Nebius Serverless AI Job. Both experiments failed the completion-quality release threshold, so this artifact is published for research and reproducibility rather than production use.
Files
adapter_model.safetensors: PEFT LoRA adapter.adapter_config.json, tokenizer files, and chat_template.jinja: adapter runtime configuration.tim-lora-f16.gguf: F16 GGUF adapter for use with a compatible Qwen3 4B base GGUF.eval/base-eval.json and eval/adapter-eval.json: retained 40-case development-suite reports.training.log: retained training log.artifact-manifest.json: reviewed SHA-256 hashes and sizes.
Training
- Base model:
Qwen/Qwen3-4B-Instruct-2507
- Method: completion-only 4-bit QLoRA
- LoRA rank and alpha: 16 and 16
- Hardware: one NVIDIA L40S GPU on Nebius
- Training examples: 702 synthetic typed-chat examples
- Training steps: 300
- Dataset generation: Qwen3.5 through DigitalOcean Serverless Inference
The canonical 702-row training package is public at immutable dataset revision
7eab2028563f17bae3a66c392d0dd9bbf1fe389f. It is the retained package used for this adapter and the package recorded by the later Nebius Serverless job.
Evaluation
On the repeatedly consulted 40-case development suite, the retained adapter reached:
- schema validity: 1.000
- action accuracy: 0.975
- expected-content accuracy: 0.6875
- premature-response rate on
wait cases: 0.000
The required expected-content threshold was 0.9500. Because the adapter did not pass that gate, these results must not be interpreted as production readiness or unbiased generalization.
Use
Load the adapter with PEFT on top of Qwen/Qwen3-4B-Instruct-2507, or use tim-lora-f16.gguf with a compatible llama.cpp base model. Outputs must be validated before downstream use.
Limitations
- The training data is synthetic.
- The 40-case suite became a development set after repeated evaluation.
- Completion quality is materially below the release threshold.
- The model can produce incorrect timing decisions or unsupported message content.
Links
Attribution
Developed and published by Jeremy Soo.
