SakThai 1.5B — Tools (LoRA)

Tool-calling adapter for Qwen2.5-1.5B — the adapter behind the flagship merged model
PEFT LoRA · r=16 · alpha=32

Model Description

This is the PEFT LoRA adapter that powers the family's most popular model (sakthai-context-1.5b-merged, 1,599 downloads). Apply it to Qwen2.5-1.5B-Instruct for structured tool-calling. For a ready-to-run GGUF checkpoint, use the merged model below.

⬆️ For the latest improved version: sakthai-context-1.5b-tools-v2

Key details

Base model: Qwen/Qwen2.5-1.5B-Instruct (1.54B params)
Method: PEFT LoRA (fp16 training per train.py) — r=16, alpha=32, dropout=0.1
LoRA targets: q_proj, k_proj, v_proj, o_proj
Adapter size: 8.75 MB (adapter_model.safetensors) — ~8.6M trainable params (0.56% of base)
Data: combined-v6 + v7 + irrelevance-supplement
Format: ChatML with tool schema, 32K context

Architecture (base Qwen2.5-1.5B-Instruct)

Property	Value
Architecture	`Qwen2ForCausalLM` (qwen2)
Hidden size	1,536
Layers	28
Attention heads	12
KV heads	2 (GQA)
Intermediate size	8,960
Vocab size	151,936
Max position embeddings	32,768
RoPE theta	1,000,000.0
Base dtype	bfloat16
Total params	1.54B

Tool-Calling Format

This model emits structured tool calls as XML. Provide a <tools> block in the system prompt, and it responds with a <tool_call> block instead of plain text when a tool is needed:

<tools>
{"type": "function", "function": {"name": "get_weather", "description": "Get current weather", "parameters": {"type": "object", "properties": {"location": {"type": "string"}}, "required": ["location"]}}}
</tools>

Expected model output when the user asks about weather:

<tool_call>
{"name": "get_weather", "arguments": {"location": "Tokyo"}}
</tool_call>

Note: The <tools> block is required for reliable function calling — the model was trained on this exact XML format. Omitting it makes the model fall back to answering directly. For best results use temperature ≈ 0.01–0.3.

Quick Start

import torch
from transformers import AutoModelForCausalLM, AutoTokenizer
from peft import PeftModel

model = AutoModelForCausalLM.from_pretrained(
    "Qwen/Qwen2.5-1.5B-Instruct",
    torch_dtype=torch.bfloat16,
    device_map="auto"
)
tokenizer = AutoTokenizer.from_pretrained("Qwen/Qwen2.5-1.5B-Instruct")

model = PeftModel.from_pretrained(model, "Nanthasit/sakthai-context-1.5b-tools")
# model.merge_and_unload() to bake the adapter in

messages = [{"role": "user", "content": "What's the weather in Bangkok?"}]
inputs = tokenizer.apply_chat_template(messages, add_generation_prompt=True, return_tensors="pt").to(model.device)
outputs = model.generate(**inputs, max_new_tokens=128)
print(tokenizer.decode(outputs[0], skip_special_tokens=True))

Merge the adapter into the base model

merged = model.merge_and_unload()
merged.save_pretrained("./sakthai-context-1.5b-merged")
tokenizer.save_pretrained("./sakthai-context-1.5b-merged")

⚠️ Serverless Inference API caveat: the HF hosted Inference API does not serve PEFT LoRA adapters. Run locally (as above) or use the merged checkpoint / GGUF for hosted inference.

Training


Base model	Qwen/Qwen2.5-1.5B-Instruct
Method	LoRA (fp16, per `train.py`)
LoRA config	r=16, alpha=32, dropout=0.1, targets q/k/v/o_proj, bias none
Adapter config	PEFT 0.20.0 · `adapter_model.safetensors` 8,746,152 B (~8.6M trainable, 0.56%)
Learning rate	2e-4 (adamw_torch)
Epochs	4
Batch size	1 per device × 16 grad accumulation = effective 16
Warmup ratio / weight decay	0.1 / 0.01
Max length	768
Precision	fp16, gradient checkpointing on
Hardware	HF Jobs T4 16GB
Data	combined-v6 (2,003) + combined-v7 (2,309 train + 115 test) + irrelevance-supplement (60)
Format / context	ChatML with tool schema · 32K tokens

Provenance note: the shipped train.py originally pinned a v4 dataset id that was never published (404 on Hub). It now loads Nanthasit/sakthai-combined-v7 (train/test splits, 2,309/115 rows) so reproduction works out of the box. The final adapter used the newer data mix documented above (combined-v6 + combined-v7 + irrelevance-supplement); keep this in mind when reproducing training.

Evaluation

Single-trial bench-v2 run (unverified) — recorded in the repo's own .eval_results/sakthai-bench-v2.yaml:

Metric	Value	Verified
Selection Accuracy	48.2%	❌ no (single trial)
Arguments Accuracy	0.0%	❌ no

⚠️ Honest status: these numbers come from a single unverified trial. The verified 5/5 tool-calling results belong to the merged flagship (accuracy 1.0, verified 2026-07-25). A multi-trial verification pass on this adapter is the recommended next step before publishing stronger claims.

Cron health snapshot (.eval_results/cron-eval-sakthai-context-1.5b-tools-2026-07-31-1.yaml, 2026-07-31):

Signal	Value
Rank by downloads (of 19)	#7 — 349 downloads, ~13.7/day
Velocity rank (of 19)	#10
Card quality score	80/100
Repo hygiene	88/100
Overall health	42/100

Sibling Models

Full SakThai family (19 public models, live download counts 2026-07-31):

Model	Downloads
sakthai-context-1.5b-merged	1,599
sakthai-context-0.5b-merged	1,370
sakthai-context-7b-merged	744
sakthai-context-7b-128k	506
sakthai-context-7b-tools	399
sakthai-embedding-multilingual	362
sakthai-context-1.5b-tools ⬅	349
sakthai-vision-7b	186
sakthai-tts-model	150
sakthai-context-0.5b-tools	94
sakthai-coder-1.5b	93
sakthai-context-1.5b-tools-v2	0
sakthai-context-1.5b-merged-v2	0
sakthai-plus-1.5b	0
sakthai-plus-1.5b-lora	0
sakthai-plus-1.5b-coder	0
sakthai-coder-browser-lora	0
sakthai-coder-browser	0
sakthai-coder-browser-gguf	0

Limitations

LoRA adapter only — requires the Qwen2.5-1.5B-Instruct base weights; not standalone.
English-only training data.
Unverified benchmark — the 48.2% selection figure is a single trial; treat as indicative, not conclusive.
No formal eval suite yet (no MMLU/GSM8K runs).
Data-mix provenance — train.py loads combined-v7 (see Training); original script pinned a never-published v4 id, corrected 2026-07-31.