LoRA-distilled variant of Qwen/Qwen3-1.7B, fine-tuned to power the cerebellum (small-brain) of Bīja — a memory-system-as-AI built on the eight-consciousness theory.

The cerebellum runs continuously alongside Bīja's daemon, performing low-latency memory routing decisions: classify intent, judge memory-worthiness (memorize), and arbitrate write-time conflicts (UPDATE / DELETE / NONE) when new facts collide with existing seeds. The base 1.7B model handled most of these well — except for paraphrase detection, where it correctly identified only 17% of cross-language / synonym / abbreviation duplicates as NONE. This adapter fixes that to 100%.

Why this model exists

Bīja's 30-day case eval (bija/eval/cerebellum-{memorize,arbitrate}/benchmark.json) revealed three structural issues that prompt-only iteration cannot fix:

Task	Baseline 1.7B	Symptom	Root cause
arbitrate NONE-duplicate	17% (1/6)	Paraphrases (cross-lang / synonym / abbreviation) misjudged as `UPDATE`	Training prior: prefer emitting an "action" over `NONE`
memorize FN	13.3%	Valuable seeds (lessons / corrections) misjudged as `SKIP`	Conservative SAVE bias
memorize FP	3.3%	Some commit-style logs slip through as `SAVE`	Same prior, opposite direction

A separate experiment with Granite 3.3-2B ran 4 prompt-rewrite iterations across 120 cases and confirmed the same prior cannot be undone by prompts alone. Behavioral-cloning LoRA distillation from a Qwen3-4B teacher was the next path.

Results

Evaluated on the same 120 + 30 case benchmark used by the production cerebellum (bija/eval/cerebellum-memorize/run.ts + bija/eval/cerebellum-arbitrate/run-with-sim.ts):

Metric	Baseline (Qwen3-1.7B-Q8_0 prompt-only)	LoRA Q8_0 GGUF	Δ
memorize accuracy	91.7% (110/120)	97.4% (excl 5 cold-start parse-fails)	+5.7pp
memorize FP rate	3.3%	3.3%	0
memorize FN rate	13.3%	1.7%	−11.6pp
memorize avg latency	480ms	436ms	−9%
arbitrate accuracy	76.7% (23/30)	86.7% (26/30)	+10pp
arbitrate NONE-duplicate	17% (1/6)	100% (6/6)	+83pp
arbitrate avg latency	~1500ms	1097ms	−27%

Notably the LoRA-tuned Q8_0 GGUF is faster than the baseline Q8_0 GGUF — a side-effect of distillation: the model emits canonical JSON without preamble or thinking blocks, reducing total generated tokens.

A more detailed comparison vs the MLX fp16 evaluation is in the project repo's Phase 5 wrap-up.

Files in this repo

File	Purpose
`Qwen3-1.7B-BIJA-cerebellum-Q8_0.gguf` (1.7 GB)	Drop-in Q8_0 GGUF; `llama.cpp` / Ollama / cerebellum-style sidecars load it directly
`adapters.safetensors` (38 MB)	Raw LoRA weights — apply on top of vanilla `Qwen/Qwen3-1.7B` (HF format) with `mlx_lm.fuse` or `peft`
`adapter_config.json`	mlx-lm LoRA config: rank=16, scale=2.0, dropout=0.05, num_layers=16, target=`q_proj+v_proj`

How to use

Drop-in replacement (recommended) — llama.cpp / Ollama

bash

hf download doncxy/BIJA-cerebellum-Qwen3-1.7B-v1 \
  Qwen3-1.7B-BIJA-cerebellum-Q8_0.gguf \
  --local-dir ~/models

llama-server -m ~/models/Qwen3-1.7B-BIJA-cerebellum-Q8_0.gguf -c 4096

Or for Bīja users — replace the production GGUF directly:

bash

mv ~/.seeddb/cerebellum/models/Qwen3-1.7B-Q8_0.gguf{,.baseline}
ln -s ~/models/Qwen3-1.7B-BIJA-cerebellum-Q8_0.gguf \
      ~/.seeddb/cerebellum/models/Qwen3-1.7B-Q8_0.gguf
pkill -f llama-server   # next call respawns sidecar with new weights

Apply LoRA on top of vanilla Qwen3-1.7B (MLX)

bash

pip install mlx-lm
hf download doncxy/BIJA-cerebellum-Qwen3-1.7B-v1 \
  adapters.safetensors adapter_config.json --local-dir ./bija-cerebellum-lora

mlx_lm.generate \
  --model Qwen/Qwen3-1.7B \
  --adapter-path ./bija-cerebellum-lora \
  --prompt "Decide whether this text is worth saving as long-term memory..." \
  --max-tokens 128

Training recipe

Field	Value
Base model	`Qwen/Qwen3-1.7B` (1.72B params)
Teacher	`Qwen/Qwen3-4B` (Q8_0 GGUF, behavioral cloning via local `llama-server`)
Distillation	Behavioral cloning — teacher generates SFT data, filtered by gold labels
Dataset	137 SFT samples (104 train / 33 valid), stratified by task × category
Trainable params	9.96M (0.579% of base)
LoRA rank / scale	16 / 2.0 (effective alpha 32)
LoRA dropout	0.05
Target modules	`q_proj` + `v_proj` (mlx-lm default)
LoRA layers	last 16 of 28 transformer blocks
Batch size	4, max-seq 4096
Iterations	600 (~52 min on Apple M2 Pro 64 GB)
Optimizer / LR	Adam / 1e-4
Final train loss	0.006
Best val loss	0.077 (iter 350); final 0.086
Peak memory	33.3 GB / 64 GB (fp16, no QLoRA / no grad checkpoint)
Tokens/sec	~820 avg

Intended use

Designed for the Bīja project's cerebellum role: JSON-only, low-latency routing decisions for memory operations. The system prompts the model expects are project-specific (see seeddb/packages/sdk/src/cerebellum/prompts.ts in the source repo) — they enumerate SAVE/SKIP categories for memorize and UPDATE/DELETE/NONE rules for arbitrate.

This is not a general-purpose chat model. Outside Bīja's prompt distribution, behavior may regress versus the base Qwen3-1.7B. For general use, prefer the base model.

Limitations

Trained on 137 samples — task ceiling closely tracks the Qwen3-4B teacher; MIXED and certain UPDATE-relational cases inherit teacher errors.
Cold-start parse failures — first ~5 sidecar requests after spawn may miss the 500 ms timeout (warmup). Persistent daemons amortize this away.
Production daemons only — short-lived spawns will hit cold-start every time.
Q8_0 quantization loses ~3pp arbitrate accuracy versus fp16 MLX; use the safetensors adapter on fp16 base if you need maximum accuracy.

Citation / acknowledgements

Built on:

Qwen/Qwen3-1.7B (base)
Qwen/Qwen3-4B (teacher; via local Q8_0 GGUF)
mlx-lm (training + fuse)
llama.cpp (HF→GGUF conversion)

License

Apache 2.0 (matches base model).