AlephFunk/bonsai-8b-prime-env-adapters

Bonsai 8B (1-bit) adapters for seven Prime Intellect Environments Hub tasks

Rank-16 LoRAs (all seven projections, 43.6M parameters, 87 MB as f16 GGUF) for PrismML's 1-bit Bonsai 8B, one per Prime Intellect Hub environment, trained from each environment's own rewards or gold on one laptop GPU. They are the adapters of the paper One gigabyte, many skills: task adapters on a 1-bit 8B model, scored on Prime Intellect environments (code, results and paper).

folder	Hub environment (metric)	training signal	1-bit base → + adapter (held-out items)
reverse-text/	primeintellect/reverse-text (LCS ratio)	800 env gold reversals, SFT	0.115 → 0.817 (200)
ascii-tree/	primeintellect/ascii-tree (tree similarity)	800 env gold trees, SFT	0.062 → 0.981 (200)
gsm8k/	primeintellect/gsm8k (accuracy)	rejection sampling, 757 correct own samples	0.851 → 0.871 (1,319)
medmcqa/	maziyar/OpenMed_MedMCQA (accuracy)	rejection sampling, 465 correct own samples	0.409 → 0.453 (1,000)
pubmedqa/	maziyar/OpenMed_PubMedQA (accuracy)	rejection sampling, 556 correct own samples	0.364 → 0.434 (500)
reasoning-core/	primeintellect/reasoning-core (task score)	800 env gold answers (no derivations), SFT	0.220 → 0.300 (500)
ifeval/	arcee-ai/ifeval (constraint pass rate)	rejection sampling, 479 passing own samples	0.640 → 0.640 (200)

For scale, the 4-bit Qwen3-8B (Q4_K_M, 5.03 GB) scores 0.208 / 0.230 / 0.933 / 0.567 / 0.432 / 0.256 / 0.685 on the same rows. Non-thinking mode throughout. The reasoning-core adapter changes answer shape rather than adding reasoning: it wins on some task families and loses on others (see §3.8 of the paper).

Each folder holds adapter_model.safetensors + adapter_config.json (PEFT, for prism-ml/Bonsai-8B-unpacked revision 376f381) and lora-f16.gguf (llama.cpp LoRA for Bonsai-8B-Q1_0.gguf, served with PrismML's llama.cpp fork), plus training.json (data counts, steps, loss curve). chat_template.jinja is the template used in training and evaluation (non-thinking).

Use with llama.cpp (one 1-bit base, several skills loaded, switch per request):

llama-server -m Bonsai-8B-Q1_0.gguf --jinja --chat-template-file chat_template.jinja \
  --lora reverse-text/lora-f16.gguf --lora gsm8k/lora-f16.gguf --lora medmcqa/lora-f16.gguf --lora ifeval/lora-f16.gguf \
  --lora ascii-tree/lora-f16.gguf --lora pubmedqa/lora-f16.gguf --lora reasoning-core/lora-f16.gguf \
  --lora-init-without-apply
# then POST /lora-adapters [{"id":0,"scale":1.0},...] to pick the active skill

Route; do not apply blindly. Transformation adapters (reverse-text, ascii-tree) learn a new behaviour and are destructive off-task (GSM8K 0.845 → 0.045 with the reverse-text adapter). In the paper a 72k-parameter recursive router picks the right adapter 99.8% of the time from full prompts (evaluated on the original four).

Notes. The adapters also load on Qwen3-8B (same architecture): on a Q4_K_M converted from the exact Qwen/Qwen3-8B weights the original four are neutral-to-positive (reverse-text 0.208 → 0.282, MedMCQA 0.560 → 0.590). Qwen's official Qwen3-8B-Q4_K_M.gguf is built from an AWQ-rescaled checkpoint and misapplies LoRAs trained on the published weights; see the paper.