The adapter is not a general chat model. It is a small policy model that proposes the next validated CKAN retrieval action in a Python-controlled agent loop.

Task

Given a user request, CKAN endpoint, previous tool observations, and compact catalog context, the model emits exactly one JSON action:

{"thought":"short decision summary","action":"tag_search","args":{"query":"Fahrrad","rows":10},"confidence":0.95}

Supported actions:

tag_search
group_list
organization_list
package_search
package_show
select_resource
finish
ask_clarification

Python remains responsible for validating actions, checking observed package/resource ids, executing CKAN API calls, retrying malformed output, and stopping the loop.

Quick Start

python

from peft import PeftModel
from transformers import AutoModelForCausalLM, AutoTokenizer

base_model = "openbmb/MiniCPM5-1B"
adapter = "build-small-hackathon/smolnalysis-ckan-retrieval-minicpm5-lora"

tokenizer = AutoTokenizer.from_pretrained(base_model, trust_remote_code=True)
model = AutoModelForCausalLM.from_pretrained(
    base_model,
    trust_remote_code=True,
    torch_dtype="auto",
    device_map="auto",
)
model = PeftModel.from_pretrained(model, adapter).eval()

messages = [
    {
        "role": "system",
        "content": (
            "You are the CKAN retrieval policy for smolnalysis. Emit strict JSON only. "
            "Output exactly one compact JSON object with keys: thought, action, args, confidence."
        ),
    },
    {
        "role": "user",
        "content": """Request: Zeig mir nützliche Daten zu Fahrrad.
Endpoint: https://opendata.muenchen.de/
Current state: No catalog tools have been run yet.
Observed packages: []
Observed resources: []
Observed tags: []
Observed groups: []
Observed organizations: []
Tool errors or empty results: []
Enough evidence: False
Desired next action type: tag_search
Required args schema: {"query":"search terms","rows":10}""",
    },
]

inputs = tokenizer.apply_chat_template(
    messages,
    tokenize=True,
    add_generation_prompt=True,
    enable_thinking=False,
    return_dict=True,
    return_tensors="pt",
).to(model.device)

outputs = model.generate(
    **inputs,
    max_new_tokens=384,
    do_sample=False,
    pad_token_id=tokenizer.eos_token_id,
)
print(tokenizer.decode(outputs[0][inputs["input_ids"].shape[-1]:], skip_special_tokens=True))

smolnalysis Space Configuration

text

SMOLNALYSIS_MINICPM_BACKEND=transformers
SMOLNALYSIS_MINICPM_TRANSFORMERS_MODEL_ID=openbmb/MiniCPM5-1B
SMOLNALYSIS_MINICPM_CKAN_RETRIEVAL_ADAPTER_REPO_ID=build-small-hackathon/smolnalysis-ckan-retrieval-minicpm5-lora
SMOLNALYSIS_MINICPM_CKAN_RETRIEVAL_TEMPERATURE=0
SMOLNALYSIS_MINICPM_MAX_NEW_TOKENS=384

Training Procedure

This adapter was trained with supervised fine-tuning using TRL + PEFT LoRA.

Base model: openbmb/MiniCPM5-1B
Training data: Munich Open Data CKAN inventory scenarios
Loss: assistant-only loss
LoRA rank: 16
LoRA alpha: 32
LoRA dropout: 0.05
Target modules: q_proj, k_proj, v_proj, o_proj, gate_proj, up_proj, down_proj

The training script used a MiniCPM-compatible training-only chat template with {% generation %} markers so that only assistant JSON actions contribute to loss.

Evaluation

Generation evaluation after the compact-target training run:

json

{
  "total": 160,
  "json_parse_rate": 1.0,
  "valid_action_rate": 1.0,
  "exact_action_match_rate": 1.0,
  "exact_action_matches": 160,
  "issue_counts": {}
}

Hand challenge set:

json

{
  "total": 8,
  "json_parse_rate": 1.0,
  "valid_action_rate": 1.0,
  "exact_action_match_rate": 1.0,
  "exact_action_matches": 8,
  "issue_counts": {}
}

Limitations

The adapter is trained for CKAN retrieval tool calling, not general conversation.
It should be used behind strict JSON parsing and action validation.
It may not generalize to every CKAN portal without additional traces from that portal.
It does not execute tools or verify dataset suitability by itself.

Framework Versions

PEFT 0.19.1
TRL 1.5.1
Transformers 5.11.0
PyTorch 2.12.0
Datasets 5.0.0
Tokenizers 0.22.2

Citation

bibtex

@software{vonwerra2020trl,
  title   = {{TRL: Transformers Reinforcement Learning}},
  author  = {von Werra, Leandro and Belkada, Younes and Tunstall, Lewis and Beeching, Edward and Thrush, Tristan and Lambert, Nathan and Huang, Shengyi and Rasul, Kashif and Gallouédec, Quentin},
  license = {Apache-2.0},
  url     = {https://github.com/huggingface/trl},
  year    = {2020}
}