v0.1.3

Fixed

• Model-agnostic inference: replaced hardcoded Qwen ChatML template with tokenizer.apply_chat_template() — local transformers backend now works with any model family
• vLLM model name bug: server model name from env var was not being passed through to API calls, causing 400 errors on providers like Groq

Added

• Pooled line classifier backend: new pooled backend for sentence-level classification
• LoRA auto-detection: transformers backend auto-detects and loads LoRA/PEFT checkpoints
• Batch extraction: extract_many() with concurrent requests for remote backends
• Encoder model: token-level line classification with mmBERT

Changed

• Inference refactor: shared _build_messages() used by both vLLM and transformers backends
• Query tag: renamed <task> to <query> in prompt formatting

Full changelog: v0.1.2...v0.1.3
PyPI: https://pypi.org/project/squeez/0.1.3/

v0.1.2

What's Changed

Fixed

• Chat template: switched from custom tokens to Qwen ChatML format (<|im_start|>/<|im_end|>) — critical for correct fine-tuning and inference
• Evaluation metrics: replaced line-number-based metrics with span-level exact match, precision/recall/F1, partial overlap, and empty accuracy
• Dataset metadata: recomputed all metadata fields from actual response content

Changed

• Training: added Unsloth support for memory-efficient LoRA training on Qwen 3.5
• Training config: tuned for A100 80GB (effective batch size 32, max_length 16384)
• Data splits: download_data.py now creates train/dev/test (dev for checkpoint selection)

Added

• Data quality: manually reviewed test split (55/436 corrected), traceback curation on train split (123/7148 corrected)
• Documentation: new data quality page documenting the full QA pipeline

Full Changelog: v0.1.1...v0.1.2

v0.1.1

What's Changed

• Optional heavy dependencies: torch, transformers, peft, datasets are no longer required for API-only usage
  • pip install squeez — lightweight, just openai + pyyaml (for vLLM/Groq/OpenAI-compatible servers)
  • pip install squeez[local] — adds torch, transformers, peft for local inference
  • pip install squeez[train] — adds datasets for training
  • pip install squeez[all] — everything

Full Changelog: v0.1.0...v0.1.1

v0.1.0 - Initial Release

squeez v0.1.0

First release of squeez — squeeze verbose LLM agent tool output down to only the relevant lines.

Highlights

• Two inference backends: local transformers and vLLM/OpenAI-compatible server (Groq, etc.)
• CLI + Python API: cat output.txt | squeez "Fix the bug" or use ToolOutputExtractor directly
• LoRA fine-tuning: train on the SWE-bench tool output dataset (7,148 train / 436 eval samples)
• Flexible configuration: CLI args > env vars > config file (squeez.yaml)
• Subcommands: squeez extract, squeez train, squeez eval, squeez pipeline
• Claude Code integration: use as a tool in your CLAUDE.md

Install

pip install squeez

Links

• PyPI: https://pypi.org/project/squeez/0.1.0/
• Dataset: https://huggingface.co/datasets/KRLabsOrg/tool-output-extraction-swebench