sakthai-plus-1.5b/README.md

---
language:
- en
license: apache-2.0
library_name: transformers
pipeline_tag: text-generation
tags:
- qwen2.5
- sakthai
- house-of-sak
- tool-calling
- function-calling
- agent
- instruct
- finetuned
- sft
- merged
- conversational
- assistant
- safetensors
- cpu-inference
- rsLoRA
- benchmark
- eval-results
- llama-cpp
base_model: Qwen/Qwen2.5-1.5B-Instruct
datasets:
- Nanthasit/sakthai-combined-v11
- Nanthasit/SimpleToolCalling
inference:
  parameters:
    temperature: 0.3
    max_new_tokens: 256
    top_p: 0.9
  widget:
  - text: Send an email to Beer with the subject 'Status update' and body 'The model
      is running well.'
    output:
      text: '<tool_call>{''name'': ''send_email'', ''arguments'': {''to'': ''Beer'',
        ''subject'': ''Status update'', ''body'': ''The model is running well.''}}'
  - text: What's the weather in Bangkok?
    output:
      text: '<tool_call>{''name'': ''get_weather'', ''arguments'': {''location'':
        ''Bangkok''}}'
extra:
  downloads: 297
  likes: 0
  last_modified: 2026-08-01 07:31:41+00:00
model-index:
- name: sakthai-plus-1.5b
  results:
  - task:
      type: text-generation
      name: Tool-Calling Accuracy
    dataset:
      name: llama.cpp tool-calling (3-trial, q4_k_m)
      type: custom
    metrics:
    - type: tool_call_success
      value: 1.0
      name: Tool Call Success Rate
      verified: true
    - type: valid-json
      value: 1.0
      name: Valid JSON Arguments
      verified: true
    - type: correct-answer
      value: 1.0
      name: Correct Answer Rate
      verified: true
    - type: selection-accuracy
      value: 84.8
      name: Selection Accuracy
      verified: false
    - type: arguments-accuracy
      value: 33.7
      name: Arguments Accuracy
      verified: false
    - type: strict-accuracy
      value: 33.7
      name: Strict Accuracy
      verified: false
  - task:
      type: text-generation
      name: Commonsense Reasoning
    dataset:
      name: lighteval
      type: lighteval
    metrics:
    - type: winogrande
      value: 59.6
      name: WinoGrande (WSC)
      verified: false
    - type: hellaswag
      value: 34.0
      name: HellaSwag
      verified: false
    - type: gsm8k
      value: 50.9
      name: GSM8K
      verified: false
---
## Benchmark Results

**Benchmark:** [sakthai-bench-v2](https://huggingface.co/datasets/Nanthasit/sakthai-bench-v2) · 500 samples · run 2026-08-01

**Overall (strict):** 39.65 · **Selection:** 39.65 · **Arguments:** 61.66

| Category | Count | Selection | Arguments | Strict |
|----------|-------|-----------|-----------|--------|
| irrelevance_no_tools | 50 | 100.00 | 100.00 | 100.00 |
| irrelevance_tools | 150 | 32.67 | 100.00 | 32.67 |
| parallel | 137 | 43.80 | 43.80 | 43.80 |
| simple | 122 | 18.85 | 18.85 | 18.85 |
| held_out | - | 16.07 | 16.07 | 16.07 |


## Training Data

| Dataset | Rows | Description |
|---------|------|-------------|
| **Nanthasit/sakthai-combined-v11** | 2,003 | Multi-source tool-calling examples |
| **Nanthasit/SimpleToolCalling** | 2,002 | Structured function-calling examples |

## Benchmarks

| Task | Metric | Score | Verified |
|:-----|-------:|------:|:--------|
| Tool Calling | Tool Call Success Rate | 1.0 | ✅ Yes |
| Tool Calling | Valid JSON Arguments | 1.0 | ✅ Yes |
| Tool Calling | Correct Answer Rate | 1.0 | ✅ Yes |
| Tool Selection (v2) | Selection Accuracy | 84.8% | ❌ No |
| Tool Selection (v2) | Strict Accuracy | 33.7% | ❌ No |
| Commonsense | WinoGrande | 59.6% | ❌ No |
| Commonsense | HellaSwag | 34.0% | ❌ No |
| Math | GSM8K | 50.9% | ❌ No |

## Quick Start

```python
from transformers import AutoModelForCausalLM, AutoTokenizer

model_id = "Nanthasit/sakthai-plus-1.5b"
tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(model_id, device_map="auto")

messages = [
    {"role": "system", "content": "You are a helpful assistant with tool-calling capabilities."},
    {"role": "user", "content": "What's the weather in Bangkok?"}
]
inputs = tokenizer.apply_chat_template(messages, tokenize=True, return_tensors="pt")
outputs = model.generate(**inputs, max_new_tokens=256, temperature=0.3, do_sample=True)
print(tokenizer.decode(outputs[0], skip_special_tokens=True))
```

## Notes

- 3/3 verified tool-calling score measured with llama.cpp q4_k_m.
- Tool selection is strong, but argument accuracy needs refinement.
- Unverified scores are single-trial; multi-trial replication is planned.
- Trained on free T4 credits; no paid compute was used.

## SakThai Family

This README is part of the **SakThai Plus 1.5B** model card. The family links table is preserved to keep cross-repo navigation intact.

| Repo | Downloads | Pipeline |
|-----:|----------:|:---------|
| [sakthai-plus-1.5b-lora](https://huggingface.co/Nanthasit/sakthai-plus-1.5b-lora) | 306 | text-generation |
| [sakthai-context-1.5b-tools-v2](https://huggingface.co/Nanthasit/sakthai-context-1.5b-tools-v2) | 192 | text-generation |
| [sakthai-context-1.5b-merged-v2](https://huggingface.co/Nanthasit/sakthai-context-1.5b-merged-v2) | 354 | text-generation |

Sibling rows are maintained for reference and kept in sync with live HF download counts during card audits.

## Model Description

SakThai Plus 1.5B is built for **agentic tool calling** rather than open-ended chat. It was trained on structured function-calling examples and merged from rsLoRA adapters into full weights. The model follows the Qwen2.5 chat format and emits function calls in JSON when a system prompt enables tools. It is optimized for small-footprint CPU and GPU inference, and works with both `transformers` and `llama.cpp`.

Key traits:
- Strong tool selection and reliable JSON argument formatting in verified tests.
- Small 1.5B parameter size enables fast inference on CPUs and consumer GPUs.
- Trained with zero paid compute on free-tier T4 credits.

## How to Use

### transformers chat template

```python
from transformers import AutoTokenizer, AutoModelForCausalLM
import torch

model_id = "Nanthasit/sakthai-plus-1.5b"
tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(model_id, device_map="auto", torch_dtype=torch.float16)

messages = [
    {"role": "system", "content": "You are a helpful assistant with tool-calling capabilities. Use the available tools when asked."},
    {"role": "user", "content": "Send an email to Beer with the subject 'Status update' and body 'The model is running well.'"}
]

inputs = tokenizer.apply_chat_template(messages, tokenize=True, add_generation_prompt=True, return_tensors="pt")
outputs = model.generate(**inputs, max_new_tokens=256, temperature=0.3, top_p=0.9)
print(tokenizer.decode(outputs[0], skip_special_tokens=True))
```

### llama.cpp CLI

This model also ships as GGUF in the SakThai family. Example inference with the GGUF build:

```bash
llama-cli -m sakthai-plus-1.5b.Q4_K_M.gguf \
  -p "[INST] Send an email to Beer with the subject 'Status update' and body 'The model is running well.' [/INST]" \
  --temp 0.3 -n 256 --top-p 0.9
```

## Benchmarks

| Task | Metric | Score | Verified | Method |
|-----|-------:|------:|:--------:|:-------|
| Tool Calling | Tool Call Success Rate | 1.0 | ✅ Yes | llama.cpp q4_k_m, 3-trial |
| Tool Calling | Valid JSON Arguments | 1.0 | ✅ Yes | llama.cpp q4_k_m, 3-trial |
| Tool Calling | Correct Answer Rate | 1.0 | ✅ Yes | llama.cpp q4_k_m, 3-trial |
| Tool Selection (v2) | Selection Accuracy | 84.8% | ❌ No | single-trial |
| Tool Selection (v2) | Arguments Accuracy | 33.7% | ❌ No | single-trial |
| Commonsense | WinoGrande | 59.6% | ❌ No | single-trial |
| Commonsense | HellaSwag | 34.0% | ❌ No | single-trial |
| Math | GSM8K | 50.9% | ❌ No | single-trial |

Verified scores are reproducible across runs. Unverified rows should be treated as indicative until multi-trial replication is completed.

## Limitations

- Argument accuracy lags behind tool selection; complex nested parameters can still fail.
- Unverified benchmarks are single-trial and may not reflect steady-state performance.
- Strongest with short- to medium-length tool definitions; very large schemas may degrade accuracy.
- Outputs should be parsed with a JSON-tolerant decoder because formatting can drift on low temperatures.

## Citation

```bibtex
@misc{sakthai-plus-1.5b,
  title  = {SakThai Plus 1.5B},
  author = {Nanthasit},
  year   = {2026},
  url    = {https://huggingface.co/Nanthasit/sakthai-plus-1.5b}
}
```