davidoneil/agent-sft-minicpm5-2b-lora

Agent SFT LoRA — MiniCPM5-2B

PEFT adapter for agentic tool-use / function-calling specialization.

• Base model: openbmb/MiniCPM5-2B
• Method: LoRA
• Trainer: Axolotl 0.21.2.dev0 (pinned commit in training repo)
• PEFT knobs (shared LoRA↔QLoRA): r=16, alpha=32, dropout=0.05, seed 42
• Sequence length: 4096 (sample packing)
• Epochs: 1

Training data (agent_sft_v1)

Unified mix of 61183 deduplicated examples (seed 42):

Source	Examples	License
UltraData-SFT-Agent-2609 Tool-Use	25000	Apache-2.0
APIGen-MT-5k	5000	CC-BY-NC-4.0
ToolACE	11297	Apache-2.0
xlam-function-calling-60k (capped)	19886	CC-BY-4.0

Multi-turn dialogues: 29343. Examples with tool signal: 50720.

License note: This adapter is released under CC-BY-NC-4.0 because the mix includes Salesforce CC-BY-NC data.

Load

from peft import PeftModel
from transformers import AutoModelForCausalLM, AutoTokenizer

base = "openbmb/MiniCPM5-2B"
adapter = "<this-repo>"  # or local path
tok = AutoTokenizer.from_pretrained(adapter, trust_remote_code=True)
model = AutoModelForCausalLM.from_pretrained(base, trust_remote_code=True)
model = PeftModel.from_pretrained(model, adapter)

Intended use

Research / non-commercial experiments on SLM agent specialization (tool calling, multi-turn trajectories). Evaluation is out of scope for this release; no benchmark scores are claimed here.

Training stack

• axolotl 0.21.2.dev0
• transformers 5.22.0
• peft 0.21.2
• torch 2.13.0+cu130

Citation

Please cite the base model and the datasets listed above if you use this adapter.