sakthai-context-7b-merged

Benchmark Results

Benchmark: sakthai-bench-v2 · 500 samples · run 2026-08-01

Overall (strict): 38.78 · Selection: 38.78 · Arguments: 45.53

Model Description

SakThai Context 7B Merged is a merged/continued checkpoint from the SakThai context family. It combines tool-use behaviour from Nanthasit/sakthai-context-7b-tools with long-context capability from Nanthasit/sakthai-context-7b-128k, continuing from Qwen/Qwen2.5-7B-Instruct.

It is intended for:

• Open-ended text generation with strong local execution
• Tool-calling and function-calling prompts
• Long-context agent-style task completion with structured outputs
• Offline CPU/edge deployment via llama.cpp / Ollama

Intended Use

• Text generation and chat for English use cases
• Tool-use workflows with <tools> XML prompt formatting
• Research on small-to-mid tool-calling models under MIT license

How to Use

from transformers import AutoTokenizer, AutoModelForCausalLM

model_id = "Nanthasit/sakthai-context-7b-merged"
tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(model_id, device_map="auto")

messages = [
  {"role": "user", "content": "List your available tools first, then find today's weather in Bangkok."}
]
prompt = tokenizer.apply_chat_template(messages, tokenize=False)
inputs = tokenizer(prompt, return_tensors="pt").to(model.device)
out = model.generate(**inputs, max_new_tokens=512)
print(tokenizer.decode(out[0], skip_special_tokens=True))

For GGUF inference:

from llama_cpp import Llama

llm = Llama(model_path="sakthai-context-7b-merged.Q4_K_M.gguf", n_ctx=32768, n_threads=8)
out = llm("<|user|>\nWhat tools do you have?\n<|assistant|>\n", max_tokens=256)
print(out["choices"][0]["text"])

Datasets

This line was trained/evaluated against SakThai tool-use corpora:

• Nanthasit/sakthai-combined-v6
• Nanthasit/sakthai-combined-v7
• Nanthasit/sakthai-combined-v8

Merged From

• Nanthasit/sakthai-context-7b-tools
• Nanthasit/sakthai-context-7b-128k

Benchmarks

Evaluation artifacts are preserved under .eval_results/ in this repository. Baseline comparisons should reference sibling cards:

• Nanthasit/sakthai-context-7b-tools
• Nanthasit/sakthai-context-7b-128k

Limitations

• Merge/continued checkpoint; verify behavior before production use.
• Tool calling performance depends on prompt formatting; prefer <tools> XML blocks.
• Long-context generations are most reliable up to ~8k–32k tokens with appropriate context windows.
• Benchmarking against held-out tools is encouraged before deployment.

Training

This checkpoint continues from prior SakThai context model merges. Datasets used earlier in the lineage include:

• Nanthasit/sakthai-combined-v6
• Nanthasit/sakthai-combined-v7
• Nanthasit/sakthai-combined-v8

See eval/ and .eval_results/ in this repo for local evaluation artifacts.

Citation

@misc{sakthai-context-7b-merged,
  title = {SakThai Context 7B Merged},
  author = {Nanthasit (Beer)},
  year = {2026},
  url = {https://huggingface.co/Nanthasit/sakthai-context-7b-merged}
}

License

MIT