| --- |
| license: other |
| license_name: microtensor-submission |
| license_link: https://www.microtensor.cloud/terms |
| base_model: ibm-granite/granite-4.0-h-micro |
| tags: |
| - microtensor |
| - ctx_compress |
| --- |
| |
| # mt_slm_ctx_compress |
| |
| A specialised model for the Microtensor `ctx_compress` arena (agent session compaction): it reads a |
| coding agent session of up to 120k tokens and writes the nine heading compaction summary. |
|
|
| | Field | Value | |
| |---|---| |
| | Base model | `ibm-granite/granite-4.0-h-micro` at `d5f01a3ea75f` (Apache 2.0) | |
| | Method | LoRA r=64 merged into bf16 | |
| | Arena | ctx_compress | |
| | Training data | sessions from the arena's own generator (`microtensor-io/microtensor-arenas`, `arenas/compress/synth.py`) under private seeds | |
| | Held out result | see training/runs | |
| |
| The root loads with `AutoModelForCausalLM`. Everything under `training/` is the record validators |
| require and the archive keeps; it is never scored. |
| |
