README.md

---
license: other
license_name: lfm1.0
license_link: https://huggingface.co/LiquidAI/LFM2-1.2B-Extract/blob/main/LICENSE
base_model: LiquidAI/LFM2-1.2B-Extract
language:
- en
tags:
- mlx
- lfm2
- fact-extraction
- structured-extraction
- on-device
- memory
pipeline_tag: text-generation
library_name: mlx
---

# experience-extractor-1.2b-v1 (MLX 8-bit)

A small, **on-device structured fact extractor** for memory engines, fine-tuned from
[`LiquidAI/LFM2-1.2B-Extract`](https://huggingface.co/LiquidAI/LFM2-1.2B-Extract) (LoRA (rank 32) fine-tune (mlx-lm)). It reads a chat transcript and emits every
storable fact as JSON in a fixed **8-field schema**:

```json
{"facts": [
  {"what": "...", "when": null, "where": null, "why": null,
   "who": ["..."], "fact_type": "world|experience",
   "entities": ["..."], "message_refs": ["id:m07"]}
]}
```

It powers the [`experience`](https://github.com/mindi-dev/experience) memory engine
(`EXPERIENCE_EXTRACTOR=lfm25`). This repo holds the **MLX 8-bit** build for fast inference on Apple Silicon via Ollama's MLX engine or `mlx-lm`.

## Evaluation (LongMemEval-cleaned "KU", content-recall)

> **Run it windowed.** Whole-transcript extraction caps a small model near 0.62; sliding a
> **5-message window** and unioning the per-window facts is the recall mechanism and the
> recommended deploy mode. Pairing the 350M + 1.2B as an **ensemble** reaches ~0.986 on KU.

| mode | recall | mean facts/row | repeat |
|---|---|---|---|
| **5-msg windowed** (recommended) | **0.958** | ~45 | high (use dedup) |
| 5-msg windowed + semantic dedup@0.6 | 0.931 | ~15 | ~0.16 (clean) |
| whole-transcript (single pass) | 0.389 — *free decoding understates; use windowing* | low | low |

> **This MLX 8-bit build, measured:** windowed **0.958** (= the GGUF), dedup@0.6 0.931. The low whole-transcript 0.389 is a free-decoding artifact, not quant loss — windowing recovers it.

## Files
- MLX 8-bit model (`config.json`, `model.safetensors`, tokenizer, chat template); 1.2 GB.

## Usage
**Ollama** (MLX engine, Apple Silicon):
```sh
ollama run hf.co/mindi-dev/experience-extractor-1.2b-v1-mlx-8bit
```
**mlx-lm**: `mlx_lm.generate --model mindi-dev/experience-extractor-1.2b-v1-mlx-8bit --prompt "<transcript>"`

For the recall numbers above, drive it **windowed** (5-msg sliding window + union + dedup) —
e.g. via the experience crate's `EXPERIENCE_EXTRACTION_WINDOW=5`. A single whole-transcript
pass under free decoding scores lower.


## Other formats
- GGUF (llama.cpp / Ollama / crate): [`mindi-dev/experience-extractor-1.2b-v1-GGUF`](https://huggingface.co/mindi-dev/experience-extractor-1.2b-v1-GGUF)

## Training

Full pipeline at [mindi-dev/experience](https://github.com/mindi-dev/experience) (`training/`).
Fine-tuned on real-distribution LongMemEval transcripts (leakage-safe; held-out KU never
trained on) with grounded teacher-generated labels.

## License

Fine-tune of [`LiquidAI/LFM2-1.2B-Extract`](https://huggingface.co/LiquidAI/LFM2-1.2B-Extract) under the **LFM Open License v1.0**.
Redistribution permitted with attribution + change notice; **commercial use by entities with
≥ US$10M revenue requires a Liquid AI commercial license** (Sec. 5). The crate code is MIT and
separate. See `NOTICE.md` and the full `LICENSE` in this repo.