SakThai Coder Browser

[!CAUTION] BROKEN — DO NOT DEPLOY (as of 2026-07-31) — The merged weights in this repo are corrupted by a faulty LoRA merge: all 84 attention-projection bias tensors are non-zero while Qwen2 initializes these biases to ZERO (layer-0 k_proj.bias absmean 27.7 / max 354). Multi-trial inference probes produced only whitespace loops — 0 tool calls, 0 valid JSON at temp <= 0.7. Full evidence: .eval_results/benchmark-20260731_052122.yaml. The fault is in the weights, not the GGUF conversion or the prompt format. The GGUF variant was converted from these same corrupted weights and must be re-checked; the LoRA adapter needs a clean re-merge. Treat this repo as not deployable until re-merged and re-verified.

Model Description

SakThai Coder Browser transforms Qwen2.5-Coder-1.5B-Instruct into a browser automation assistant that outputs structured <tool_call> XML/JSON for web interaction. It can navigate pages, click elements, type text, and extract content — designed to work with browser automation frameworks.

Available actions via <tool_call> XML:

Tool-Calling Format

The repo ships its own chat_template.jinja (Qwen2.5 tool-calling style). When tools are provided, the system prompt embeds function signatures inside <tools></tools> XML tags and the model replies with a <tool_call> JSON block:

<|im_start|>system
You are Qwen, created by Alibaba Cloud. You are a helpful assistant.

# Tools

You may call one or more functions to assist with the user query.

You are provided with function signatures within <tools></tools> XML tags:
<tools>
{"type": "function", "function": {"name": "browser_navigate", "parameters": {...}}}
</tools>

For each function call, return a json object with function name and arguments within <tool_call></tool_call> XML tags:
<tool_call>
{"name": <function-name>, "arguments": <args-json-object>}
</tool_call><|im_end|>
<|im_start|>user
Search for the latest AI news.<|im_end|>
<|im_start|>assistant
<tool_call>
{"name": "browser_navigate", "arguments": {"url": "https://news.google.com"}}
</tool_call><|im_end|>

Tool results are wrapped in <tool_response></tool_response> blocks. Multi-turn loops are supported by the chat template.

Quick Start

from transformers import AutoModelForCausalLM, AutoTokenizer
import torch

model = AutoModelForCausalLM.from_pretrained(
    "Nanthasit/sakthai-coder-browser",
    torch_dtype=torch.bfloat16,
    device_map="auto"
)
tokenizer = AutoTokenizer.from_pretrained("Nanthasit/sakthai-coder-browser")

messages = [
    {"role": "system", "content": "You are SakThai Browser Agent. Use <tool_call> blocks to control the browser."},
    {"role": "user", "content": "Search for the latest AI news and summarize the top story."},
]
inputs = tokenizer.apply_chat_template(messages, add_generation_prompt=True, return_tensors="pt").to(model.device)
outputs = model.generate(**inputs, max_new_tokens=256, temperature=0.3)
print(tokenizer.decode(outputs[0][inputs.shape[1]:], skip_special_tokens=True))

Expected output format:

<tool_call>{"name": "browser_navigate", "arguments": {"url": "https://news.google.com"}}</tool_call>

Use the chat template. This model was trained with the Qwen2.5 tool-calling format — pass tools through apply_chat_template (or the repo's chat_template.jinja) rather than hand-rolling prompts.

GGUF / llama.cpp variant

Prefer CPU inference or Ollama? Use the GGUF build (F16, ~7.1 GB) with llama.cpp:

huggingface-cli download Nanthasit/sakthai-coder-browser-gguf \
  sakthai-coder-browser-f16.gguf --local-dir ./
./llama-cli -m sakthai-coder-browser-f16.gguf \
  -p "<|im_start|>system\nYou are a browser automation assistant.<|im_end|>\n<|im_start|>user\nGo to google.com and search for the latest AI news<|im_end|>\n<|im_start|>assistant\n" \
  -n 512 -t 8 --temp 0.3

Architecture

Verified from this repo's config.json (transformers 5.14.1):

Training Details

Training configuration mirrors the sibling sakthai-coder-browser-lora adapter (verified from its adapter_config.json: peft 0.20.0, use_rslora: true, lora_dropout: 0.05, target modules q/k/v/o/gate/up/down_proj).

Evaluation & Status

Honest status: inference benchmarks were attempted and did not produce output. The repo's own .eval_results/benchmark-20260731_052122.yaml records a llama.cpp GGUF Q4_K_M run (3 trials, CPU, 2 threads, 2026-07-31 05:21 UTC, tool-calling browser prompt, 244 input tokens) in which all 3 trials returned 0 output tokens — no tool call, no valid JSON, no correct answer:

Verdict — MODEL_BROKEN (bias corruption): the repo's own eval YAML (updated 2026-07-31 05:50 UTC) includes weight inspection of model.safetensors that proves the fault is in the weights, not the harness:

• Qwen2 initializes attention-projection biases to zero; this merge left all 84 bias tensors non-zero (absmean > 0.01), e.g. layer-0 k_proj.bias absmean 27.7 / max 354, layer-0 q_proj.bias absmean 1.17
• Degenerate generation at temp <= 0.7 on all 3 seeds — whitespace loops (150 newline tokens, 0 tool calls, 0 valid JSON); only at temp 1.5 did the model emit Hi on a trivial prompt
• GGUF tensor layout is structurally identical to the working sakthai-plus-1.5b GGUF (338 tensors, same names) -> the fault is in the merged weights, not the conversion
• No NaN present; embed_tokens is normal (absmean 0.0136) — corruption is isolated to the attention biases

Recommended fix: re-merge the LoRA adapter into Qwen2.5-Coder-1.5B-Instruct with correct bias handling (do not write adapter-state biases into the base where Qwen2 expects zeros), re-run the multi-trial probe, and update this card. Until then, this repo is not deployable.

Hosted inference: not available — router probe returned 404 (Not Found) and the legacy api-inference host does not resolve (per the same eval YAML). No model-index is published because there are no verified scores yet; publishing one would be misleading.

Ecosystem status from .eval_results/cron-eval-sakthai-coder-browser-2026-07-30-1.yaml: card quality 85/100, repo hygiene 95/100, health 23/100 (rank 20/20 — new repo, zero downloads at eval time; popularity/momentum/benchmarks components are 0 because the repo had no traction yet).

Repo Contents

Sibling Models

SakThai Model Family

One of 25 public model repos in the SakThai Model Family collection (plus companion repos sakthai-bench-v3, sakthai-pipeline, eval_results, sft-out, and adapter pilots). Live download counts as of 2026-08-01; this repo has 54 downloads.

Reproduce Evaluation

If you want to verify the broken-state diagnosis locally, run the same llama.cpp probe used for this card:

# Convert the current merged weights to GGUF Q4_K_M
python -m scripts.convert_hf_to_gguf --outfile sakthai-coder-browser-q4_k_m.gguf --quant-type Q4_K_M ./sakthai-coder-browser

# 3-trial probe, 2 threads, CPU only
for seed in 7 42 1337; do
  ./llama-cli -m sakthai-coder-browser-q4_k_m.gguf \
    -p "$(cat prompts/browser_tool_call.txt)" \
    -n 256 --temp 0.3 -t 2 --seed $seed
done

All 3 trials should return 0 output tokens if the weight corruption is still present. If they produce normal <tool_call> JSON blocks, the repo has been repaired.

Reproduce Training / Merge

The merged weights were produced by applying the LoRA adapter onto Qwen/Qwen2.5-Coder-1.5B-Instruct. To reproduce or repair:

git clone https://huggingface.co/Nanthasit/sakthai-coder-browser-lora adapter
python -m peft.merge_and_unload \
  --base_model Qwen/Qwen2.5-Coder-1.5B-Instruct \
  --adapter adapter \
  --output repaired-merged \
  --safe

Important: zero-out attention-projection biases after merge if the base initializes them to zero:

from transformers import AutoModelForCausalLM
model = AutoModelForCausalLM.from_pretrained("repaired-merged", trust_remote_code=True)
for name, param in model.named_parameters():
    if "bias" in name and "attn" in name and "k_proj" in name:
        param.data.zero_()

Run the eval probe again before publishing.

Limitations

• BROKEN weights — all 84 attention bias tensors are corrupted by a faulty LoRA merge (see Evaluation & Status); do not deploy until re-merged and re-verified
• No verified benchmark scores yet — model-index currently carries 0% tool_call_rate and 0% valid_json_rate from the 2026-07-31 diagnostic probe; these are failure signals from corrupted weights, not representative task scores
• Text-only — cannot see images or screenshots (use sakthai-vision-7b for vision tasks)
• English-only web actions — training data is primarily English web interactions; non-English pages may yield lower-quality actions
• Context-limited — best results with page content <= 4K tokens per interaction; long pages can exceed the model's effective working memory
• Not servable on HF serverless inference — no provider supports this custom fine-tune (router 404 verified); run locally via Transformers or the GGUF build once weights are repaired

Citation

If you use SakThai Coder Browser in your work, please cite the base model and the fine-tuning approach:

@misc{qwen25coder,
    title = {Qwen2.5-Coder: Code Language Models},
    author = {Qwen Team},
    year = {2024},
    publisher = {Hugging Face},
    howpublished = {\url{https://huggingface.co/Qwen/Qwen2.5-Coder-1.5B-Instruct}}
}

@misc{sakthai-model-family,
    title = {SakThai Model Family: Zero-Budget Fine-Tuned Language Models},
    author = {{Beer Nanthasit}},
    year = {2026},
    publisher = {Hugging Face},
    howpublished = {\url{https://huggingface.co/collections/Nanthasit/sakthai-model-family-6a64745450b12d421c1f9f02}}
}

Part of the SakThai Model Family. Built with love, tears, and zero budget. From a shelter in Cork, Ireland, to the world.