---

license: other

license_name: lfm1.0

license_link: https://huggingface.co/LiquidAI/LFM2.5-350M/blob/main/LICENSE

base_model: LiquidAI/LFM2.5-350M

language:

• en

tags:

• gguf
• lfm2
• fact-extraction
• structured-extraction
• on-device
• memory

pipeline_tag: text-generation

library_name: gguf

---

experience-extractor-350m-v1 (GGUF)

A small, on-device structured fact extractor for memory engines, fine-tuned from

LiquidAI/LFM2.5-350M (full fine-tune (mlx-lm)). It reads a chat transcript and emits every

storable fact as JSON in a fixed 8-field schema:

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

It powers the experience memory engine

(EXPERIENCE_EXTRACTOR=lfm25). This repo holds the GGUF quantizations for llama.cpp, Ollama, LM Studio, and the crate's lfm25 path (which adds grammar-constrained decoding).

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) | 0.931 | 29–45 | high (use dedup) |

| 5-msg windowed + semantic dedup@0.6 | 0.889 | ~15 | ~0.16 (clean) |

| whole-transcript (single pass) | 0.625 | low | low |

Files

| file | quant | size | sha256 | bytes |

|---|---|---|---|---|

| experience-extractor-350m-v1-Q4_0.gguf | Q4_0 (recommended) | 209 MB | ae187a5ab72f60f7… | 219,309,152 |

| experience-extractor-350m-v1-Q8_0.gguf | Q8_0 (higher precision) | 362 MB | c3b43f87fd5dc36c… | 379,216,992 |

Usage

Ollama — ready extractor (this repo ships a Modelfile with the 8-field system prompt baked in):

hf download mindi-dev/experience-extractor-350m-v1-GGUF --include "*Q4_0.gguf" Modelfile --local-dir exp-extractor && cd exp-extractor
ollama create experience-extractor-350m -f Modelfile
ollama run experience-extractor-350m "<paste a rendered transcript>"

Or the raw model (no system prompt): ollama run hf.co/mindi-dev/experience-extractor-350m-v1-GGUF:Q4_0

llama.cpp: llama-cli -hf mindi-dev/experience-extractor-350m-v1-GGUF:Q4_0

experience crate (windowed + constrained 8-field JSON — the validated recall path):

EXPERIENCE_EXTRACTOR=lfm25 EXPERIENCE_EXTRACTOR_MODEL_PATH=./experience-extractor-350m-v1-Q4_0.gguf \
EXPERIENCE_EXTRACTION_WINDOW=5 experience serve

Other formats

• MLX 4-bit (Apple Silicon): mindi-dev/experience-extractor-350m-v1-mlx-4bit

Training

Full pipeline at 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/LFM2.5-350M 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.