docs: CometAPI infrastructure block
Browse files
README.md
CHANGED
|
@@ -1,64 +1,3 @@
|
|
| 1 |
-
---
|
| 2 |
-
license: apache-2.0
|
| 3 |
-
base_model: chopratejas/kompress-v2-base
|
| 4 |
-
tags:
|
| 5 |
-
- token-classification
|
| 6 |
-
- compression
|
| 7 |
-
- context-compression
|
| 8 |
-
- headroom
|
| 9 |
-
- glm
|
| 10 |
-
- synthetic-data
|
| 11 |
-
language:
|
| 12 |
-
- en
|
| 13 |
-
---
|
| 14 |
-
|
| 15 |
-
# kompress-v13 — GLM Agent Scenarios + Regex Teacher (experimental)
|
| 16 |
-
|
| 17 |
-
Training data generated by [GLM-5.1-FP8](https://huggingface.co/zai-org/GLM-5.1-FP8) simulating realistic multi-turn coding sessions across Rust, Python, TypeScript, and Go. Labels: deterministic must-keep regex (same as production safety net). **Experimental — [kompress-v8](https://huggingface.co/PeetPedro/kompress-v8) is the production recommendation.**
|
| 18 |
-
|
| 19 |
-
## Results
|
| 20 |
-
|
| 21 |
-
| Metric | v8 (domain data) | v13 (GLM scenarios) |
|
| 22 |
-
|---|---|---|
|
| 23 |
-
| heretic exact (32p) | **0.955** | 0.951 |
|
| 24 |
-
| keep_rate | 0.854 | 0.951 |
|
| 25 |
-
| override_delta | 0.000 | 0.000 |
|
| 26 |
-
|
| 27 |
-
**Finding:** GLM-generated scenarios are realistic but the regex teacher labels too many tokens as must-keep (paths, identifiers, numbers all match in code). The regex-as-teacher approach produces conservative models.
|
| 28 |
-
|
| 29 |
-
## Training
|
| 30 |
-
|
| 31 |
-
127 GLM-generated tool outputs + 254 generic (33% GLM ratio). 3 epochs from v2-base.
|
| 32 |
-
|
| 33 |
-
## Usage
|
| 34 |
-
|
| 35 |
-
```python
|
| 36 |
-
from headroom import compress, CompressConfig
|
| 37 |
-
result = compress(messages, config=CompressConfig(kompress_model="PeetPedro/kompress-v13"))
|
| 38 |
-
```
|
| 39 |
-
|
| 40 |
-
|
| 41 |
-
## CONCLUSION
|
| 42 |
-
|
| 43 |
-
Regex teacher produces conservative models. 0.951 heretic but keep_rate 0.951.
|
| 44 |
-
|
| 45 |
-
## USECASE
|
| 46 |
-
|
| 47 |
-
Demonstrates regex-as-teacher limitation. Educational value.
|
| 48 |
-
|
| 49 |
-
## Series
|
| 50 |
-
|
| 51 |
-
| Version | Teacher | Heretic | Status |
|
| 52 |
-
|---|---|---|---|
|
| 53 |
-
| v2-base | — | 0.975 | ceiling |
|
| 54 |
-
| v4 | self-labels | 0.943 | — |
|
| 55 |
-
| **v8** | **Qwen2.5-7B** | **0.955** | **production** |
|
| 56 |
-
| v13 | regex | 0.951 | too conservative |
|
| 57 |
-
|
| 58 |
-
[Full benchmark →](https://pocoo.vaked.dev/posts/2026-06-25-kompress-heretic-eval) | [Training repo →](https://github.com/peterlodri-sec/ultrawhale) | [Headroom →](https://github.com/headroomlabs-ai/headroom) | [vaked.dev →](https://pocoo.vaked.dev)
|
| 59 |
-
|
| 60 |
-
---
|
| 61 |
-
|
| 62 |
## Infrastructure
|
| 63 |
|
| 64 |
Model API provider: CometAPI
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
## Infrastructure
|
| 2 |
|
| 3 |
Model API provider: CometAPI
|