You need to agree to share your Hugging Face information to access this model
This repository is publicly accessible, but you have to accept the conditions to access its files and content.
Fine-tuned on web-agent trajectories built on public benchmark tasks (see the dataset card for per-source licenses). Access is reviewed manually. Please describe your intended use.
Log in or Sign Up to review the conditions and access this model content.
Minion web agent β v7d
LoRA adapter on lukakuma/minion_l32_1d2B_l30_grpo (revision 3e208e836e416b17719d74ed81eef4bf54f9db26), trained to act as every LLM call of a DOM-based,
tool-calling web agent (actions, planning tools, goal decomposition, completion check) from the agent's own
successful trajectories.
Training
| item | value |
|---|---|
| dataset | lukakuma/web-trajectories-sft config cap3 at 53a08ba00497 |
| context window | 131072 tokens (llama3 RoPE factor 16.0) |
| loss | assistant outputs except turns whose tool call failed or repeated the previous call |
| packing | consecutive calls of a session in one sequence, hints visible only to their own turn |
| LoRA rank / alpha | 16 / 32 |
| learning rate / epochs / warmup | 0.0001 / 2 / 0.03 |
| micro-batches per step | 1 |
| optimizer steps | 1445 |
| training documents / target tokens | 9272 / 6894683 |
Evaluation
| λͺ¨λΈ | μ¨λΌμΈ μ±κ³΅ (95% ꡬκ°) | μ | νμ μ ν¨ | μ€ν€λ§ | λꡬ μ΄λ¦ | μΈμ μ ν | νλ / κ³ν / λ³΄κ³ λꡬ μ΄λ¦ | IFEval | GSM8K |
|---|---|---|---|---|---|---|---|---|---|
| minion_grpo | β | β | β | β | β | β | β / β / β | β | β |
| v7b | β | β | 99.5% | 99.4% | 47.0% | 13.8% | 46.5% / 52.6% / 41.2% | 49.4% | 66.9% |
| v7d | β | β | 99.6% | 99.6% | 48.7% | 14.6% | 48.6% / 51.9% / 43.5% | 50.8% | 67.2% |
μ¨λΌμΈ: λΌμ΄ λ κ³Όμ (lab/benchmarks/public/slices/web_trajectories_test_v1.json)λ₯Ό μ€μ λΈλΌμ°μ λ‘ μ€νν΄ λ°μ΄ν°μ κ°μ νμ κΈ°λ‘ μ±μ ν μ
μ±κ³΅λ₯ . μ€νλΌμΈ: test νλ³Έ(λꡬ νΈμΆ n=1897)μμ μλ μμ΄μ νΈμ λ€μ νΈμΆκ³Ό λΉκ΅ β μ λ΅μ΄ νλκ° μλλ―λ‘ λ³΄μ‘° μ§ν. IFEval prompt-level strict, GSM8K flexible-extract.
Use
from peft import PeftModel
from transformers import AutoModelForCausalLM, AutoTokenizer
base = AutoModelForCausalLM.from_pretrained("lukakuma/minion_l32_1d2B_l30_grpo", revision="3e208e836e416b17719d74ed81eef4bf54f9db26")
model = PeftModel.from_pretrained(base, "lukakuma/minion-agent-lora-v7d").merge_and_unload()
tokenizer = AutoTokenizer.from_pretrained("lukakuma/minion_l32_1d2B_l30_grpo", revision="3e208e836e416b17719d74ed81eef4bf54f9db26")
The adapter was trained with the window above: set rope_parameters / max_position_embeddings of the base config
to the values in model-config.json before loading, or use the merged weights.
The merged weights are in merged/: vllm serve lukakuma/minion-agent-lora-v7d --revision main --tokenizer lukakuma/minion_l32_1d2B_l30_grpo ... after downloading that folder, or load them with AutoModelForCausalLM.from_pretrained("lukakuma/minion-agent-lora-v7d", subfolder="merged").
Tool calls are emitted as <tool_call>{"name": ..., "arguments": {...}}</tool_call> (vLLM: --tool-call-parser hermes).
- Downloads last month
- -
Model tree for lukakuma/minion-agent-lora-v7d
Base model
lukakuma/minion_l32_1d2B_l30_grpo