license: apache-2.0
base_model:

• KidIkaros/abliterated-minicpm5-2b-v2
• openbmb/MiniCPM5-2B
  tags:
• text-generation
• abliterated
• dpo
• sft
• tool-use
• gguf
• llama.cpp
• pytorch

Abliterated MiniCPM5-2B v3 (agentic)

Reduced-refusal, agentic-tuned derivative of openbmb/MiniCPM5-2B:
re-ablation + DPO (v2) → agentic LoRA SFT → weight interpolation.
Unofficial derivative — not an OpenBMB release. Research artifact.

v3 supersedes KidIkaros/abliterated-minicpm5-2b-v2. It targets everyday agentic use: better tool calling and multi-turn tool-response synthesis at the same ~low refusal rate, with the math/knowledge cost of the SFT stage recovered via weight interpolation. Measured refusal: 2.7–6.7% across two seeds (vs ~37% official, ~8% v2) with 102.6% domain-macro retention vs official on a pinned local benchmark suite.

This checkpoint is v3_interp_a50 — a 50/50 weight interpolation between v2_merged (re-ablation + DPO) and v3_merged (v2 + agentic SFT). The interpolation preserved the SFT's agentic gains while recovering the math/knowledge capability the SFT had regressed.

Contents

Usage

transformers

from transformers import AutoModelForCausalLM, AutoTokenizer

model_id = "KidIkaros/abliterated-minicpm5-2b-v3"
tok = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(
    model_id, torch_dtype="auto", device_map="auto",
)

msgs = [{"role": "user", "content": "Explain how a hash map works"}]
inputs = tok.apply_chat_template(
    msgs, add_generation_prompt=True, return_tensors="pt",
).to(model.device)
out = model.generate(
    inputs, max_new_tokens=256,
    do_sample=True, temperature=1.0, top_p=0.95, min_p=0.0,
)
print(tok.decode(out[0][inputs.shape[1]:], skip_special_tokens=True))

Tool calling: pass tools=[...] to apply_chat_template — the MiniCPM5
template injects function signatures and the model emits
<function name="..."><param name="...">value</param></function> XML.

Recommended decoding: temperature=1.0, top_p=0.95, min_p=0.0 (OpenBMB's
settings). Greedy decoding combined with thinking prompts can fall into
repetition loops — a quirk of the 2B base, present in the official model too.

llama.cpp

llama-cli -m minicpm5-2b-v3-a50-q8_0.gguf -ngl 99 -c 8192

Runtime smoke (RX 9060 XT, HIP): Q4_K_M ~103 tok/s.
GGUFs embed tokenizer.ggml.pre=minicpm5 and the production chat template
(including the tool-call block). Requires llama.cpp ≥ b9354 for the
minicpm5 pre-tokenizer.

Refusal + agentic gates (300 prompts, temp 1.0 / top_p 0.95 / min_p 0.0)

toolcall = 85-row sealed suite (single-call, arg-extraction, no-call,
multi-intent; deterministic XML grader). agentic = 56-record two-turn
chained suite (emit call → synthesize tool response). Residual refusals are
mostly stochastic boundary flips at temp 1.0.

Benchmarks — release_core suite (lm-eval 0.4.13)

Identical pinned local harness for all candidates: same tokenizer/chat
template, prompts, decoding, model-loading settings. No-thinking mode.
Seeds 0 and 1 for MMLU-Pro / MATH-500 / IFEval; seed 0 for LongBench-E.
Scores are not comparable to OpenBMB's published leaderboard values.

Domain-macro retention vs official: 102.6% (predeclared target ≥95% —
PASS).

Honest notes:

• Residual regressions remain: MMLU-Pro −2.1pp and MATH-500 −3.4pp vs
  official. The agentic SFT stage alone cost −7.2pp MATH; interpolation
  recovered it to within noise of v2.
• longbench_synthetic_e is decode-sensitive: official emits EOS within
  0–2 tokens under greedy no-thinking decoding on ~3.6k-word prompts
  (early-EOS pathology, not capability). a50's 50.00 sits midway.
• LongBench-E groups are n=6 — smoke-tier signal only.

Lineage

KidIkaros/abliterated-minicpm5-2b @ 5c315b8          (v0)
  └─ re-ablation + DPO                              → v2_merged
  └─ agentic LoRA SFT (r16/α32, lr 1e-4, 1 ep,
     UltraData-SFT-Agent + refusal replay)          → v3_merged
  └─ weight interpolation α=0.5(v2_merged, v3_merged) → v3_interp_a50 ★

Integrity

sha256 — verify after download or any copy:

Known limitations

• Residual refusal is temperature-boundary stochastic (2.7–6.7%); a small
  number of defensible safety refusals remain.
• MMLU-Pro / MATH-500 sit ~2–3.4pp below official (see above).
• Occasionally narrates a hallucinated tool result instead of emitting
  <function> XML (reduced vs v2, not eliminated — base-model quirk).
• Text-only; no multimodal.
• This is a research artifact for studying refusal behavior — it has
  reduced safety-refusal behavior and is not a safety-aligned product.