Modelpublic

Pseudogiant-xr/pseudolife-extractor-gemma-4-e4b

sourceHugging Facegemmaupdated 2mo agoView on Hugging Face

2likes93downloads

Model Card

PseudoLife Extractor E4B

A task-specialized QLoRA fine-tune of Gemma 4 E4B (instruction-tuned) that distills agent session transcripts into structured long-term memory: canonical facts (entity / attribute / value), typed relations, and procedural lessons. It is the default "dream consolidation" extractor for PseudoLife-MCP, a persistent neural-memory system for LLM agents (public code release forthcoming).

Shipped as a merged, Q4_K_M-quantized GGUF (~5.3 GB) that runs CPU-only under llama.cpp at roughly 12–15 tok/s on a desktop CPU — no GPU required.

Files

File	Version	Notes
`pseudolife-extractor-e4b-v2-Q4_K_M.gguf`	v2	Current. Registry-datagen fine-tune (see below).

Version-suffixed filenames are stable: pin the exact file in download URLs. Future extractor versions will be added to this same repository as -v3, -v4, ….

Training

v2 was trained with QLoRA on 887 distilled extraction examples generated by a Claude Sonnet teacher using a per-chain key registry datagen scheme: the teacher is required to re-state carried-forward facts under their previously minted keys each session, which teaches the student consistent key reuse instead of key proliferation. The LoRA was merged into the base weights before quantization.

Evaluation

Scored on a LongMemEval-derived knowledge-update benchmark (same-session comparison against the previous PseudoLife extractor fine-tune of the same base model):

Metric	v2 (this model)	prior fine-tune
KU-oracle, cortex (fact-spine) accuracy	0.705	0.603
KU-oracle, hybrid retrieval accuracy	0.756	0.744
Extraction ladder: gold facts recovered	0.9	1.0
Extraction ladder: stale-fact leakage	0.0	0.0

Usage

Serve with llama.cpp (the daemon-facing OpenAI-compatible endpoint PseudoLife-MCP expects):

bash

llama-server -m pseudolife-extractor-e4b-v2-Q4_K_M.gguf --port 8081 -c 8192

Recommended sampling is embedded in the GGUF metadata (temperature 1.0, top_k 64, top_p 0.95).

Scope note: this is a narrow specialist, not a chat model. It is trained against PseudoLife-MCP's combined facts + relations extraction prompt format and will underperform on general instruction following. In particular, the E4B base engages poorly with relations-only prompts — always request facts and relations together.

License and provenance

This model is a derivative of Gemma and is provided under and subject to the Gemma Terms of Use. By downloading or using this model you agree to those terms, including the Gemma Prohibited Use Policy. Downstream distributors must pass these terms on to their users.

Fine-tune, quantization, and evaluation by Pseudogiant-xr.