entelligentsia/grove-explore-base-GGUF

grove-explore-base

Untrained Qwen3.5-4B base, self-converted + quantized, adopted as the grove explore locator delegate.

Untrained base. This is the off-the-shelf upstream Qwen3.5-4B (Alibaba / Qwen), self-converted to GGUF and quantized, adopted as the locator delegate for grove's delegated exploration mode (grove init --as mcp-llm). It is not a grove fine-tune. Full lineage: PROVENANCE.md.

Quants

Quant	Size	file_type	Eval (sheet-coverage)	ollama tag	Role
Q8_0	4.6 GB	7	82.1 (n=347)	q8_0	canonical baseline
Q4_K_M	2.78 GB	15	80.6 (n=347)	q4_k_m	interim winner — memory-lean serving default

Default / recommended: Q4_K_M (q4_k_m) — best size/quality trade-off. The higher-precision quant is the canonical eval baseline. Full metrics below.

Evaluation

Measured on the grove explore holdout (347 episodes across 9 pinned real-world repos spanning 9 languages), via the is-grep-enough bench (hardened judge, 2026-07-14). These are grove's own delegation metrics, not a public leaderboard benchmark.

Quant	Answer-sheet coverage	Grounding gold rate	No-answer	Holdout
Q8_0	82.1	89% (308/347)	38	347
Q4_K_M	80.6	83% (288/347)	56	347

• sheet_coverage — mean answer-sheet text-match coverage, 0-100, higher is better (completeness of the located answer).
• grounding_gold_pct — share of episodes whose citations pass the grounding filter (valid, evidenced file:line) — higher is better.
• no_answer — episodes that produced no citation at all — lower is better.

Grounding gold rate by repository

Share of episodes with valid, evidenced citations, per pinned repo (9 real-world codebases, one per language).

Repo	Language	Q8_0	Q4_K_M
bitcoin	C++	97%	81%
django	Python	92%	89%
hugo	Go	89%	96%
laravel	PHP	86%	67%
rails	Ruby	95%	95%
redis	C	97%	82%
spring-boot	Java	95%	88%
typescript	TypeScript	74%	66%
webpack	JavaScript	76%	82%

Apples-to-apples caveat: the two rows are the same weights at different precision — Q8_0 is the canonical baseline; Q4_K_M is the memory-lean default. Grounding policy: min_fs=0.8, min_evidence=0.5.

Using it with grove

grove's inner explorer drives a tool-calling loop against this model over an OpenAI-compatible endpoint and returns validated file:line citations. Point .grove/explore.json at your local endpoint:

{
  "provider": "ollama",
  "base_url": "http://localhost:11434/v1",
  "model": "grove-explore-base:q4_k_m",
  "steering": "strict"
}

ollama (requires ollama ≥ 0.32)

ollama pull entelligentsia/grove-explore-base:q4_k_m
# or import a local gguf with the shipped recipe:
#   ollama create grove-explore-base:q4_k_m -f Modelfile.q4_k_m

llama.cpp / llama-server

llama-server -hf entelligentsia/grove-explore-base-GGUF:Q4_K_M \
    --alias grove-explore-base --jinja -ngl 99 \
    -c 98304 -np 4 --cache-type-k q8_0 --cache-type-v q8_0

--alias grove-explore-base makes the served /v1/models id stable regardless of file path; --jinja uses the embedded chat template (chat_template.jinja).

Serving notes

• Thinking must be ON. Serve with thinking ON — thinking off yields degenerate empty tool-calls. ollama returns chain-of-thought in a separate reasoning field, the answer in content.
• Recommended num_ctx for grove: 24576; temperature: 0.
• Architecture: qwen35, 441 tensors, 4-section rope [11, 11, 10, 0], context
  262144. Self-converted specifically to preserve the 4-section rope (3-section GGUFs from other converters will not load in standalone llama.cpp).

Integrity (sha256)

• 54ea292c6551c5608e1d014ca2d71ec6432314c18ffba81e61bce36bf1af042a grove-explore-base-q8_0.gguf
• 1062c26f69f0aba645a3b8eeaeebfbe0c2fbc15c2f7fab284f437ef3f0391056 grove-explore-base-q4_k_m.gguf

License

apache-2.0, inherited from the upstream Qwen3.5-4B weights this redistributes (upstream LICENSE). Full text + attribution: LICENSE.