Tara 1.3 is a tiny experimental AI-engineering tool-call model. It is trained to read a short User: prompt and emit a compact JSON object that either selects a tool or answers with tool: "none".

This release is best treated as a research checkpoint for structured tool-routing experiments, not as a production assistant.

Model Details

Model name: tara1.3
Internal checkpoint: tara-1.3-ai-engineer-toolcall-sft-v4-plaintok-from300/checkpoint-100
Architecture: LlamaForCausalLM
Context length: 1,024 tokens
Vocabulary size: 16,384
Hidden size: 512
Layers: 7
Attention heads: 8
Weights format: safetensors
License: Apache-2.0

Capability

Tara 1.3 is designed to produce JSON-style tool calls for a small AI-engineering tool set.

Supported tool names:

weather
search
segment
evaluate_model
train_sft
inspect_file
extract_json
none

Example target shapes:

{"tool":"weather","location":"Bangkok tomorrow"}

{"tool":"search","query":"Python list comprehension examples"}

{"tool":"train_sft","base_model":"models/base","dataset":"data/train.txt","output_dir":"outputs/run"}

{"tool":"none","response":"Tokenizer validation passed. Next, run a small generation smoke test."}

Quick Start

python

import json
import torch
from transformers import AutoModelForCausalLM, AutoTokenizer

repo_id = "aungkomyint/tara1.3"

tokenizer = AutoTokenizer.from_pretrained(repo_id)
model = AutoModelForCausalLM.from_pretrained(repo_id)
model.eval()

def generate_tool_call(user_text):
    prompt = f"User: {user_text.strip()}\nAssistant:\n"
    inputs = tokenizer(prompt, return_tensors="pt")
    inputs.pop("token_type_ids", None)

    with torch.no_grad():
        output = model.generate(
            **inputs,
            max_new_tokens=120,
            do_sample=False,
            repetition_penalty=1.08,
            pad_token_id=tokenizer.pad_token_id,
            eos_token_id=tokenizer.eos_token_id,
        )

    text = tokenizer.decode(output[0], skip_special_tokens=False)
    reply = text[len(prompt):] if text.startswith(prompt) else text.split("Assistant:", 1)[-1]
    reply = reply.split("<|endoftext|>", 1)[0].split("<|pad|>", 1)[0].strip()
    return reply

reply = generate_tool_call("Search the web for Python list comprehension examples.")
print(reply)

try:
    parsed = json.loads(reply)
    print("tool:", parsed.get("tool"))
except json.JSONDecodeError:
    print("Model did not return valid JSON for this sample.")

Recommended Prompt Format

Use a simple instruction format:

text

User: <request>
Assistant:

Greedy decoding is recommended for tool-call tests:

text

do_sample=False
max_new_tokens=120
repetition_penalty=1.08

Training Summary

Tara 1.3 was trained as a supervised fine-tuning continuation for AI-engineering tool calls.

Training configuration:

Steps: 300
Block size: 1,024
Batch size: 8
Gradient accumulation: 4
Effective batch size: 32
Learning rate: 5e-5
Warmup steps: 15
Weight decay: 0.01
Loss mask: only the final Assistant response is trained; earlier turns are context

Dataset:

Train examples: 8,945
Eval examples: 777
Mixture: no-tool/chat behavior plus capped tool-call examples

Local Evaluation

The released checkpoint was selected from a small local comparison on 2026-06-24.

The 10-prompt eval covered weather, search, segmentation, model evaluation, SFT training, file inspection, JSON extraction, and no-tool/general responses.

Checkpoint	Valid JSON	Schema OK	Expected Tool Match
`checkpoint-100`	6/10	5/10	5/10
`checkpoint-200`	5/10	5/10	5/10
`checkpoint-300`	5/10	5/10	5/10

checkpoint-100 was selected because it tied the other continued checkpoints on schema correctness and tool selection while producing one more valid JSON output.

Tokenizer validation passed: tool-call JSON tokenizes through the plain BPE vocabulary without old chat/tool special tokens.

Limitations

This is a very small experimental model.
It can emit malformed JSON.
It can choose the right tool but fill arguments with copied or unrelated values.
General tool: "none" responses are unstable.
It is not reliable for autonomous tool execution without validation, repair, and fallback logic.
It should not be used for medical, legal, financial, safety, or other high-stakes decisions.

Suggested Runtime Guardrails

Applications should:

Parse the output with a JSON parser.
Validate the tool name against an allowlist.
Validate required fields for each tool.
Reject or repair malformed JSON.
Require user confirmation before destructive or external actions.

Citation

If you use this model, cite it as:

Aung Ko Myint. Tara 1.3. 2026. Hugging Face model checkpoint.