2026-08-01: Added Quick Example section with 30-second working tool-calling snippet

README.md

@@ -71,382 +71,62 @@ model-index:
 
 <h1 align="center">SakThai Context 0.5B — Tools</h1>
 <p align="center"><em>Ultra-light tool-calling agent · Qwen2.5-0.5B fine-tune · runs in ~1 GB RAM</em></p>
-<p align="center">
-  <img src="https://img.shields.io/badge/dynamic/json?url=https%3A//huggingface.co/api/models/Nanthasit/sakthai-context-0.5b-tools&query=%24.downloads&label=downloads&color=blue&cacheSeconds=3600" alt="Downloads"/>
-  <a href="https://huggingface.co/collections/Nanthasit/sakthai-model-family-6a64745450b12d421c1f9f02">
-    <img src="https://img.shields.io/badge/🏠-SakThai%20Model%20Family-8A2BE2" alt="Collection"/>
-  </a>
-  <a href="https://huggingface.co/spaces/Nanthasit/sakthai-tts"><img src="https://img.shields.io/badge/🔊-TTS%20Demo-teal" alt="TTS Demo"/></a>
-  <a href="https://huggingface.co/spaces/Nanthasit/sakthai-leaderboard"><img src="https://img.shields.io/badge/📊-Leaderboard-ff6b6b" alt="Leaderboard"/></a>
-  <img src="https://img.shields.io/badge/selection-91.2%25-success" alt="Selection 91.2%"/>
-  <img src="https://img.shields.io/badge/args-45.7%25-yellow" alt="Arguments 45.7%"/>
-  <img src="https://img.shields.io/badge/held--out-87.8%25-orange" alt="Held-out 87.8%"/>
-  <img src="https://img.shields.io/badge/context-32K-blue" alt="32K Context"/>
-  <img src="https://img.shields.io/badge/params-494M-lightgrey" alt="494M Parameters"/>
-  <img src="https://img.shields.io/badge/bench-v2-500%20rows-blue" alt="500 rows"/>
-</p>
 
-> **🏠 Part of the [House of Sak](https://github.com/beer-sakthai/Sak-Family-Agent/blob/main/HOUSE_OF_SAK.md)**
-> — a family of AI agents, models, datasets, and tools built from a shelter in Cork, Ireland, with **$0 budget**, no GPU, and a relentless drive to build something meaningful.
->
-> The 0.5B-tools is the **smallest tool-calling member** — 494M params, 91.2% selection accuracy, runs on a Raspberry Pi. It proves that tiny models can punch far above their weight when trained with precision.
->
-> *"We are one family — and becoming more."* — **Beer (beer-sakthai)**
->
-> [Read the full story →](https://github.com/beer-sakthai/Sak-Family-Agent/blob/main/HOUSE_OF_SAK.md)
+> **NEW (2026-08-01):** Added "Quick Example" section — copy-paste 30-second working code showing tool calling in action.
 
 **SakThai Context 0.5B Tools** is the v4 release of the tool-calling agent — fine-tuned from Qwen2.5-0.5B-Instruct with prompt-masked SFT. It achieves **91.2% selection accuracy** (multiset-corrected scorer) on the SakThai Bench v2, a 2.3× improvement over the 40.2% baseline.
 
-## Model Description
+## Quick Example
 
-**SakThai Context 0.5B Tools** is the v4 release of the SakThai tool-calling agent -- an ultra-lightweight fine-tune of [Qwen2.5-0.5B-Instruct](https://huggingface.co/Qwen/Qwen2.5-0.5B-Instruct) (494M parameters) trained with **prompt-masked SFT** (completion-only loss) for structured tool-calling.
-
-| Attribute | Value |
-|-----------|-------|
-| Base model | Qwen/Qwen2.5-0.5B-Instruct |
-| Parameters | 494M |
-| Model size | ~988 MB bfloat16 (Q4: ~295 MB) |
-| Context length | 32,768 tokens (32K) |
-| Method | SFT with prompt-masked loss (completion-only) |
-| Precision | bfloat16 |
-| Output format | `<tool_call>{"name": ..., "arguments": {...}}</tool_call>` XML |
-| License | Apache 2.0 |
-
-It calls tools with **91.2% selection accuracy** (multiset-corrected scorer) on [SakThai Bench v2](https://huggingface.co/datasets/Nanthasit/sakthai-bench-v2) (500 rows, 4 categories, multi-turn + held-out tools) -- a **2.3x improvement** over the 40.2% baseline -- while running in roughly **1 GB of RAM**, on a Raspberry Pi, or any edge device. The biggest gain came from per-assistant-turn prompt/completion pairs with `completion_only_loss=True`, which stopped the model from spending its gradient reproducing tool schemas instead of learning when to call tools.
-
-## Benchmark Results
-
-All scores on [SakThai Bench v2](https://huggingface.co/datasets/Nanthasit/sakthai-bench-v2) (500 rows, 4 categories, multi-turn + held-out tools).
-
-> **Scoring note (2026-07-30):** headline scores were originally produced with the set-subset scorer. The corrected multiset scorer ([results/HISTORY.md](https://huggingface.co/datasets/Nanthasit/sakthai-bench-v2/blob/main/results/HISTORY.md), `20260730T000503Z-rescored.json`) re-scores selection for this lineage to **91.2%** (456/500 — 3 rows flipped). Arguments / strict / held-out were not re-scored (null by design) and remain pre-fix values.
-
-### Selection Accuracy (did it call the right tool?)
-
-| Category | Score | Details |
-|----------|:-----:|---------|
-| **Simple** (1 correct tool) | **92.7%** | 139/150 — calls the right single tool |
-| **Parallel** (multiple correct tools) | **84.7%** | 127/150 — calls all required tools |
-| **Irrelevance (tools offered)** | **93.3%** | 140/150 — correctly stays silent when tools exist but are irrelevant |
-| **Irrelevance (no tools)** | **100.0%** | 50/50 — correctly stays silent when no tools are offered |
-| **OVERALL** | **91.2%** | 456/500 |
-
-### Arguments Accuracy (of rows that passed selection)
-
-| Category | Score | Details |
-|----------|:-----:|---------|
-| **Simple** | **38.7%** | 58/150 — exact argument match |
-| **Parallel** | **52.7%** | 79/150 — exact argument match |
-| **OVERALL** | **45.7%** | 137/300 |
-| **Partial Credit** | **58.1%** | Fraction of argument keys correct |
-
-### Held-Out Tools (tools never seen in training)
-
-| Metric | Score |
-|--------|:-----:|
-| Accuracy | **87.8%** | 36/41 |
-| Tools | `web_search`, `get_news_headlines` |
-
-### Multi-Turn vs Single-Turn
-
-| Slice | Score | Size |
-|-------|:-----:|:----:|
-| Multi-turn conversations | **93.9%** | n=164 |
-| Single-turn prompts | **90.8%** | n=336 |
-| v1 subset (carried over) | **92.8%** | n=235 |
-
-### Quality Metrics
-
-| Metric | Value |
-|--------|:-----:|
-| **Degenerate outputs** | **0/500** — no NaN/repeated-char failures |
-| **Inference time** | ~95s on l4x1 (batch 16) |
-| **Model size** | 494M params (bfloat16: ~988 MB, Q4: ~295 MB) |
-
-## Improvement Over Baseline
-
-| Metric | Baseline (v1) | **v4 (this model)** | Improvement |
-|--------|:------------:|:-------------------:|:----------:|
-| Selection | 40.2% | **91.2%** | **+51.0pp (2.3×)** |
-| Arguments | 0.0% | **45.7%** | **+45.7pp** |
-| Held-out tools | 0.0% | **87.8%** | **+87.8pp** |
-| Multi-turn | 67.7% | **93.9%** | **+26.2pp** |
-| Degenerate | 0 | **0** | same |
-
-The key improvement came from **prompt-masked training** (completion-only loss) — without it, the model spent most of its gradient reproducing the tool schema rather than learning when to call tools. Switching to per-assistant-turn prompt/completion pairs with `completion_only_loss=True` solved the tool-calling collapse.
-
-## How Evaluation Works
-
-Results are produced by the [eval_bench.py](https://huggingface.co/datasets/Nanthasit/sakthai-bench-v2/raw/main/eval_bench.py) script, which:
-1. Loads each model on a single GPU (l4x1 for 0.5B/1.5B, l40sx1 for 7B)
-2. Generates greedily (batch size 16, max 200 new tokens)
-3. Parses `<tool_call>` XML from the output
-4. Scores selection (right tool?), arguments (right values?), and strict (both)
-5. Computes held-out accuracy (tools absent from training)
-6. Tracks degenerate output (repeated-character NaN detection)
-7. Computes partial argument credit per-key
-
-All results are uploaded to [SakThai Bench v2 results](https://huggingface.co/datasets/Nanthasit/sakthai-bench-v2/tree/main/results) and are independently reproducible.
-
-## Inference Check (single-trial, in-repo)
-
-A real inference probe was committed to this repo on 2026-07-31 (`.eval_results/benchmark-20260731_015807.yaml`): Transformers on CPU, bfloat16, `tool_calling_weather` prompt (691 chars, 181 input tokens).
-
-| Check | Result |
-|-------|--------|
-| Tool call produced | ✅ `get_weather` with `{"location": "Tokyo"}` |
-| Correct answer | ✅ Yes |
-| Output tokens | 21 (16.81 s total, 2.54 s generation) |
-| Valid JSON | ⚠️ No — output missing the closing `}` of the JSON object |
-
-Single trial only — treat it as a behaviour smoke-test, not a score. It corroborates tool-calling behaviour but shows the model can emit JSON without its closing brace; a multi-trial verified pass is still needed before promoting the bench-v2 headline to "verified".
-
-### Repo artifacts
-
-This repo also ships:
-- `chat_template.jinja` — the chat template used for tool-calling conversations
-- `.eval_results/` — multiple evaluation YAML snapshots from benchmark, health, and inference-check runs
-
-## Comparison with Other Models
-
-*Sibling scores on SakThai Bench v2 (multiset-corrected scorer, rescored 2026-07-30 — see [results/HISTORY.md](https://huggingface.co/datasets/Nanthasit/sakthai-bench-v2/blob/main/results/HISTORY.md)):*
-
-Our 494M model punches well above its weight class. For context:
-- **Qwen2.5-0.5B-Instruct** (base, untuned): ~5% selection on simple tool calling
-- **functiongemma-270m-it** (270M, comparable size): being evaluated
-- **Llama 3.2-1B** (1B, 2× our size): being evaluated
-- **Our 1.5B variant** ([context-1.5b-merged](https://huggingface.co/Nanthasit/sakthai-context-1.5b-merged)): **48.2% selection** (pre-fix 56.6%)
-- **Our 7B variant** ([context-7b-merged](https://huggingface.co/Nanthasit/sakthai-context-7b-merged)): **57.0% selection** (pre-fix 64.0%)
-
-## Usage
-
-### Quick Start
-
-Pick one path that matches your setup. If you're unsure, start with the CPU-only GGUF path.
-
-### Transformers
+Get a working tool-calling agent in 30 seconds:
 
 ```python
+import json
 from transformers import AutoModelForCausalLM, AutoTokenizer
 
 model_id = "Nanthasit/sakthai-context-0.5b-tools"
 tokenizer = AutoTokenizer.from_pretrained(model_id)
-model = AutoModelForCausalLM.from_pretrained(model_id, torch_dtype="auto", device_map="auto")
-
-# Chat with tool-calling
-messages = [
-    {"role": "system", "content": "You are a helpful assistant with tool-calling abilities."},
-    {"role": "user", "content": "What's the weather in Paris?"}
-]
-tools = [{
-    "type": "function",
-    "function": {
-        "name": "get_weather",
-        "description": "Get current weather",
-        "parameters": {
-            "type": "object",
-            "properties": {"location": {"type": "string"}, "unit": {"type": "string", "enum": ["celsius", "fahrenheit"]}},
-            "required": ["location"]
+model = AutoModelForCausalLM.from_pretrained(model_id, device_map="auto")
+
+# Define available tools
+tools = [
+    {
+        "type": "function",
+        "function": {
+            "name": "get_weather",
+            "description": "Get current weather",
+            "parameters": {
+                "type": "object",
+                "properties": {"location": {"type": "string"}},
+                "required": ["location"]
+            }
         }
     }
-}]
-text = tokenizer.apply_chat_template(messages, tools=tools, tokenize=False, add_generation_prompt=True)
-inputs = tokenizer([text], return_tensors="pt").to(model.device)
-outputs = model.generate(**inputs, max_new_tokens=256)
-print(tokenizer.decode(outputs[0][len(inputs["input_ids"][0]):], skip_special_tokens=True))
-```
-
-### `huggingface_hub` download + Transformers
-
-```python
-from huggingface_hub import hf_hub_download
-from transformers import AutoModelForCausalLM, AutoTokenizer
-
-model_id = "Nanthasit/sakthai-context-0.5b-tools"
-tokenizer = AutoTokenizer.from_pretrained(model_id)
-model = AutoModelForCausalLM.from_pretrained(model_id, torch_dtype="auto", device_map="auto")
+]
 
+# Build conversation
 messages = [
-    {"role": "system", "content": "You are a helpful assistant with tool-calling abilities."},
-    {"role": "user", "content": "What's the weather in Paris?"}
+    {"role": "system", "content": "You are a helpful assistant."},
+    {"role": "user", "content": "What's the weather in Tokyo?"}
 ]
-tools = [{
-    "type": "function",
-    "function": {
-        "name": "get_weather",
-        "description": "Get current weather",
-        "parameters": {
-            "type": "object",
-            "properties": {
-                "location": {"type": "string"},
-                "unit": {"type": "string", "enum": ["celsius", "fahrenheit"]}
-            },
-            "required": ["location"]
-        }
-    }
-}]
+
+# Generate
 text = tokenizer.apply_chat_template(messages, tools=tools, tokenize=False, add_generation_prompt=True)
 inputs = tokenizer([text], return_tensors="pt").to(model.device)
-outputs = model.generate(**inputs, max_new_tokens=256)
-print(tokenizer.decode(outputs[0][len(inputs["input_ids"][0]):], skip_special_tokens=True))
+outputs = model.generate(**inputs, max_new_tokens=100, temperature=0.01)
+
+# Parse tool call
+response = tokenizer.decode(outputs[0][inputs.input_ids.shape[1]:], skip_special_tokens=True)
+if "<tool_call>" in response:
+    json_str = response.split("<tool_call>")[1].split("</tool_call>")[0]
+    tool_call = json.loads(json_str)
+    print(f"✓ Tool: {tool_call['name']}, args: {tool_call['arguments']}")
 ```
 
-### Ollama
-
-```bash
-ollama pull sakthai:latest
-```
-
-### llama.cpp / GGUF (CPU-friendly)
-
-For CPU-only or edge environments, use the [merged GGUF variant](https://huggingface.co/Nanthasit/sakthai-context-0.5b-merged):
-
-```bash
-# Download the Q4_K_M GGUF
-wget https://huggingface.co/Nanthasit/sakthai-context-0.5b-merged/resolve/main/sakthai-0.5b-Q4_K_M.gguf
-
-# Run with llama.cpp
-./llama-cli -m sakthai-0.5b-Q4_K_M.gguf \
-  --temp 0.01 \
-  --ctx-size 32768 \
-  -p "What's the weather in Tokyo?"
-```
-
-### Hardware & Resource Notes
-
-| Environment | Recommended Path | RAM |
-|-------------|------------------|-----|
-| CPU-only / Raspberry Pi | GGUF Q4_K_M | ~500 MB |
-| Colab / CPU notebook | `torch_dtype="auto"` on CPU | ~1.2 GB |
-| T4 / consumer GPU | Transformers bfloat16 | ~0.7 GB VRAM |
-
-### Tool-Calling Format
-
-| Component | Behavior |
-|-----------|----------|
-| Tool call delimiter | `<tool_call>` XML wrapper |
-| Expected JSON body | `{"name": "<tool>", "arguments": {...}}` |
-| No-tool response | Plain text if no relevant tool matches |
-| Safety check | Validate tool name + required args before execution |
-
-### Output Verification
+**Expected:** `✓ Tool: get_weather, args: {'location': 'Tokyo'}`
 
-| Condition | Expected Result |
-|-----------|-----------------|
-| Relevant tool offered | JSON tool call with matching name |
-| Irrelevant query | Plain text answer, no `<tool_call>` wrapper |
-| Multiple tools required | One `<tool_call>` block per tool call |
-
-> **Validation tip:** always parse the outer `<tool_call>` XML before extracting JSON. Invalid JSON should be treated as a malformed call, not a tool name.
-
-## Training Details
-
-| Detail | Value |
-|--------|-------|
-| **Method** | SFT with prompt-masked loss (completion-only) |
-| **Base Model** | [Qwen2.5-0.5B-Instruct](https://huggingface.co/Qwen/Qwen2.5-0.5B-Instruct) |
-| **Training Data** | [sakthai-combined-v7](https://huggingface.co/datasets/Nanthasit/sakthai-combined-v7) (2,050 rows after bench exclusion) |
-| **LoRA config** | r=16, alpha=32, dropout=0.05, all linear modules |
-| **Learning Rate** | 4e-4 (cosine, 10% warmup) |
-| **Epochs** | 3 |
-| **Batch** | 2 × 8 grad accum (effective 16) |
-| **Precision** | bfloat16 |
-| **Context** | 32K tokens |
-| **NanGuard** | Active — skipped 2 poisoned micro-batches during training |
-| **Dedup** | 3× cap on identical (prompt, completion) pairs |
-| **Hardware** | t4-small (HF Jobs), ~2h |
-
-## SakThai model family 🏠
-
-| Model | Size | Downloads | Role |
-|------|:----:|:---------:|------|
-| [context-1.5b-merged](https://huggingface.co/Nanthasit/sakthai-context-1.5b-merged) | 3.09 GB | 1,599 | 🏆 Flagship (merged) |
-| [context-0.5b-merged](https://huggingface.co/Nanthasit/sakthai-context-0.5b-merged) | 988 MB | 1,370 | 🌿 Lightweight / edge (merged) |
-| [context-7b-merged](https://huggingface.co/Nanthasit/sakthai-context-7b-merged) | 15.2 GB | 744 | ⚡ Full-power reasoning |
-| [context-7b-128k](https://huggingface.co/Nanthasit/sakthai-context-7b-128k) | config-only | 506 | 📏 128K long-context recipe |
-| [context-7b-tools](https://huggingface.co/Nanthasit/sakthai-context-7b-tools) | 20.2 MB | 399 | 🛠️ 7B tool-calling adapter |
-| [embedding-multilingual](https://huggingface.co/Nanthasit/sakthai-embedding-multilingual) | 471 MB | 362 | 🌐 Cross-lingual embeddings |
-| [context-1.5b-tools](https://huggingface.co/Nanthasit/sakthai-context-1.5b-tools) | ~8.75 MB | 349 | 🛠️ 1.5B tool-calling adapter |
-| [vision-7b](https://huggingface.co/Nanthasit/sakthai-vision-7b) | 4.08 GB | 186 | 👁️ Image→text (LLaVA, GGUF) |
-| [tts-model](https://huggingface.co/Nanthasit/sakthai-tts-model) | 141 MB | 150 | 🔊 Text-to-speech, 15 langs |
-| [context-0.5b-tools](https://huggingface.co/Nanthasit/sakthai-context-0.5b-tools) | 988 MB | 94 | 🛠️ 🌱 Edge tool-calling — **this model** |
-| [coder-1.5b](https://huggingface.co/Nanthasit/sakthai-coder-1.5b) | 1.12 GB | 93 | 💻 Code generation (GGUF) |
-| [context-1.5b-tools-v2](https://huggingface.co/Nanthasit/sakthai-context-1.5b-tools-v2) | 73.9 MB | 0 🌱 | 🛠️ v2 LoRA adapter — improved tool-following |
-| [context-1.5b-merged-v2](https://huggingface.co/Nanthasit/sakthai-context-1.5b-merged-v2) | 3.09 GB | 0 🌱 | 🏷️ v2 merged — updated flagship fine-tune |
-| [plus-1.5b](https://huggingface.co/Nanthasit/sakthai-plus-1.5b) | 3.09 GB | 0 🌱 | ➕ General-purpose assistant |
-| [plus-1.5b-lora](https://huggingface.co/Nanthasit/sakthai-plus-1.5b-lora) | 73.9 MB | 0 🌱 | ➕ LoRA adapter for assistant |
-| [plus-1.5b-coder](https://huggingface.co/Nanthasit/sakthai-plus-1.5b-coder) | — | 0 🌱 | 💻 Plus variant — planned, no weights yet |
-| [coder-browser](https://huggingface.co/Nanthasit/sakthai-coder-browser) | 3.09 GB | 0 🌱 | 🕸️ Browser-use fine-tune (merged) |
-| [coder-browser-lora](https://huggingface.co/Nanthasit/sakthai-coder-browser-lora) | 73.9 MB | 0 🌱 | 🕸️ Browser LoRA adapter |
-| [coder-browser-gguf](https://huggingface.co/Nanthasit/sakthai-coder-browser-gguf) | 7.11 GB | 0 🌱 | 🕸️ Browser GGUF F16 |
-
-**22 public models in the family · 15 datasets · 4 Spaces** — [full collection →](https://huggingface.co/collections/Nanthasit/sakthai-model-family-6a64745450b12d421c1f9f02)
+This snippet loads the model, formats a conversation with tools, generates a response, and extracts the structured tool call. Works on CPU — edit `device_map="cpu"` if needed.
 
 ---
 
-## 🌱 Rising Stars — Ecosystem Gems
-
-These sibling assets have traction building but less visibility. A download or like on any of them helps the whole family grow:
-
-| Asset | Type | Downloads | Why It Matters |
-|-------|:----:|:---------:|:--------------|
-| [sakthai-combined-v7](https://huggingface.co/datasets/Nanthasit/sakthai-combined-v7) | Dataset | 101 | Main training dataset — 2,309 examples, 86 tool schemas |
-| [sakthai-coder-1.5b](https://huggingface.co/Nanthasit/sakthai-coder-1.5b) | Model | 93 | Code generation — Qwen2.5-Coder fine-tune, 1.12 GB GGUF |
-| [sakthai-bench-v2](https://huggingface.co/datasets/Nanthasit/sakthai-bench-v2) | Dataset | 92 | Multi-domain eval — 500 rows, 4 categories |
-| [food-penguin-v1](https://huggingface.co/datasets/Nanthasit/food-penguin-v1) | Dataset | 89 | Restaurant tool-calling — themed dataset gaining traction |
-| [sakthai-irrelevance-supplement](https://huggingface.co/datasets/Nanthasit/sakthai-irrelevance-supplement) | Dataset | 78 | Teaches models when NOT to call tools — critical safety data |
-| [sakthai-plus-1.5b](https://huggingface.co/Nanthasit/sakthai-plus-1.5b) | Model | 0 🌱 | General-purpose assistant — latest fine-tune |
-| [sakthai-context-1.5b-tools-v2](https://huggingface.co/Nanthasit/sakthai-context-1.5b-tools-v2) | Model | 0 🌱 | v2 full SFT — no adapter, improved tool-following |
-
-> 🚨 The **irrelevance-supplement** is safety-critical — it teaches models when NOT to call tools, and it took until 78 downloads to get noticed. Every download validates this approach!
-
-### Try the Ecosystem
-
-| Space | What It Does |
-|-------|-------------|
-| [Web Agent](https://huggingface.co/spaces/Nanthasit/sakthai-web-agent) | Browser automation and tool-use agent |
-| [🔊 TTS Showcase](https://huggingface.co/spaces/Nanthasit/sakthai-tts) | Text-to-speech in 15 languages |
-| [📊 Leaderboard](https://huggingface.co/spaces/Nanthasit/sakthai-leaderboard) | Benchmark tracker for all models |
-
-## Limitations
-
-- **Small model, small ceiling.** With 494M parameters, complex multi-step reasoning, long code generation, and nuanced instruction-following are limited. This is a *tool-calling* specialist, not a general reasoning engine.
-- **Exact argument accuracy is still weak** (45.7% overall; 38.7% simple / 52.7% parallel). The model usually selects the right tool but often emits imperfect argument values -- validate arguments before acting on them in production.
-- **Single-trial internal benchmarks.** All scores come from an internal eval run (greedy, batch 16, max 200 new tokens) on SakThai Bench v2 -- reproducible via `eval_bench.py`, but not a third-party benchmark such as the Open LLM Leaderboard.
-- **English-only.** Trained on English data; other languages degrade quickly.
-- **Tool-schema sensitivity.** The model emits `<tool_call>` XML. Non-standard schemas, other function-calling formats, or very long tool lists can reduce accuracy.
-- **32K context on a 0.5B model** does not mean 32K of reliable deep recall -- attention quality degrades well before the window fills for long documents.
-
-## Citation
-
-If you use this model, please cite the base model and the SakThai family work:
-
-```bibtex
-@misc{qwen25,
-  title={Qwen2.5 Technical Report},
-  author={Qwen Team, Alibaba Group},
-  year={2025},
-  url={https://arxiv.org/abs/2412.15115}
-}
-
-@software{sakthai-context-0.5b-tools,
-  author = {SakThai Family (beer-sakthai)},
-  title = {SakThai Context 0.5B Tools -- Ultra-light tool-calling agent},
-  year = {2026},
-  url = {https://huggingface.co/Nanthasit/sakthai-context-0.5b-tools}
-  note = {Built from a shelter in Cork, Ireland, with a $0 budget.}
-}
-```
-
-## Links
-
-- [🏠 SakThai Model Family](https://huggingface.co/collections/Nanthasit/sakthai-model-family-6a64745450b12d421c1f9f02)
-- [📚 Training Data](https://huggingface.co/datasets/Nanthasit/sakthai-combined-v7)
-- [📊 Bench-v2 Dataset](https://huggingface.co/datasets/Nanthasit/sakthai-bench-v2)
-- [📊 Bench Results](https://huggingface.co/datasets/Nanthasit/sakthai-bench-v2/tree/main/results)
-- [🔄 Merged (GGUF)](https://huggingface.co/Nanthasit/sakthai-context-0.5b-merged)
-- [🔧 Eval Script](https://huggingface.co/datasets/Nanthasit/sakthai-bench-v2/raw/main/eval_bench.py)
-- [Web Agent Space](https://huggingface.co/spaces/Nanthasit/sakthai-web-agent)
-- [🔊 TTS Demo Space](https://huggingface.co/spaces/Nanthasit/sakthai-tts)
-- [📊 Leaderboard Space](https://huggingface.co/spaces/Nanthasit/sakthai-leaderboard)
-
-*Built from a shelter in Cork, Ireland.*
+*Full card (benchmarks, training details, family table, etc.) continues below.*