TaimoorSiddiqui

Hopcoder-Mini-9B-SWE-Agent-LoRA-H200

Model Details

Table

Base model	`TaimoorSiddiqui/Hopcoder-Mini-9B`
Stage-1 adapter	`TaimoorSiddiqui/Hopcoder-Mini-9B-Native-ToolCall-LoRA-H200`
Adapter type	LoRA (r=16, alpha=32)
Training hardware	NVIDIA H200 (Modal)
Training time	~3h 47m
Total steps	659 (1 epoch)
Eval loss	0.4701 (step 200)
Learning rate	2.0e-5
Batch size	2 × grad_accum 8 = 16 effective
Max sequence length	4096

Training Data

Table with columns: Source, Examples
Source	Examples
Open-SWE patches	5,489
xLAM function-calling	1,200
Hermes function-calling	341
Targeted CLI tool-call replay	3,840
Total	10,870

Native Tool-Call Format

The model uses a custom XML-like format for tool calls (not JSON):

markdown
<function=tool_name>
<parameter=param_name>
param_value
</parameter>
</function>

Multiple tool calls are separated by newlines. This format was introduced in Stage-1 and reinforced in Stage-2.

Benchmark Results

Both benchmarks were run on a Modal H200 container loading the base model + adapter with bfloat16 precision.

1. Native Tool-Call Benchmark (44 cases)

The native benchmark evaluates 40 tool-use cases (5 per tool across 8 tools) and 4 no-tool cases.

Table with columns: Metric, Score
Metric	Score
Correct tool selection	79.5%
Required args present	88.6%
Native format valid	100.0%
Correct call count	88.6%
JSON params valid	100.0%
No old JSON format	100.0%
No markdown fences	100.0%
Balanced tags	100.0%

Per-Tool Accuracy

Table with columns: Tool, Accuracy, Args Present, Correct/Total
Tool	Accuracy	Args Present	Correct/Total
`ask_user_question`	100.0%	100.0%	5/5
`edit`	100.0%	100.0%	5/5
`glob`	100.0%	100.0%	5/5

2. BFCL-Compatible Evaluation (10 cases)

The BFCL-style evaluation tests general function-calling with unseen tools (weather, flights, email, stocks, etc.).

Table with columns: Metric, Score
Metric	Score
Correct tool selection	100.0%
Native format valid	100.0%
No old JSON format	100.0%

All 10 BFCL test cases passed with perfect tool selection and format compliance.

Key Findings

Perfect format compliance — 100% native XML format, zero legacy JSON, zero markdown fences, zero extra prose. The Stage-2 adapter fully internalized the native tool-call format.
Strong general tool selection — 6 of 8 tools scored 100% accuracy. ask_user_question, edit, glob, read_file, run_shell_command, and todo_write are fully reliable.
Generalization to unseen tools — The 100% BFCL score confirms the model generalizes its tool-calling ability to functions it has never seen during training.
Weaknesses identified:
- search_code (40%) — the model sometimes substitutes grep_search instead of search_code, or calls with -style arguments.

Usage

python
from peft import PeftModel
from transformers import AutoModelForCausalLM, AutoTokenizer

base_model = "TaimoorSiddiqui/Hopcoder-Mini-9B"
adapter_repo = "TaimoorSiddiqui/Hopcoder-Mini-9B-SWE-Agent-LoRA-H200"

tokenizer = AutoTokenizer.from_pretrained(base_model)
model = AutoModelForCausalLM.from_pretrained(base_model, torch_dtype="auto", device_map="auto")
model = PeftModel.from_pretrained(model, adapter_repo)

Files

Table with columns: File, Description
File	Description
`adapter_config.json`	LoRA configuration
`adapter_model.safetensors`	Adapter weights
`stage2_benchmark_results.json`	Combined benchmark results (native + BFCL)
`stage2_native_benchmark.json`	Full native tool-call benchmark with per-case details
`stage2_bfcl_eval.json`	Full BFCL evaluation with per-case details

Citation

bibtex
@misc{hopcoder-mini-9b-swe-agent-lora-h200,
  author = {TaimoorSiddiqui},
  title = {HopCoder-Mini-9B SWE-Agent LoRA H200},
  year = {2026},
  url = {https://huggingface.co/TaimoorSiddiqui/Hopcoder-Mini-9B-SWE-Agent-LoRA-H200}
}

License

This adapter inherits the license of the base model TaimoorSiddiqui/Hopcoder-Mini-9B.

Available on FriendliAI

Dedicated Endpoints

Run this model inference on single tenant GPU with unmatched speed and reliability at scale.

Learn more

Model Details

Model Provider

TaimoorSiddiqui

Model Tree

Base

TaimoorSiddiqui/Hopcoder-Mini-9B

Adapter

this model

Input Modalities

TextImageVideo

Output Modalities

Text

Supported Functionality

Dedicated Endpoints

Explore FriendliAI today

Get started Talk to an engineer

Model Details

Table

Base model	`TaimoorSiddiqui/Hopcoder-Mini-9B`
Stage-1 adapter	`TaimoorSiddiqui/Hopcoder-Mini-9B-Native-ToolCall-LoRA-H200`
Adapter type	LoRA (r=16, alpha=32)
Training hardware	NVIDIA H200 (Modal)
Training time	~3h 47m
Total steps	659 (1 epoch)
Eval loss	0.4701 (step 200)
Learning rate	2.0e-5
Batch size	2 × grad_accum 8 = 16 effective
Max sequence length	4096

Training Data

Table with columns: Source, Examples
Source	Examples
Open-SWE patches	5,489
xLAM function-calling	1,200
Hermes function-calling	341
Targeted CLI tool-call replay	3,840
Total	10,870

Native Tool-Call Format

The model uses a custom XML-like format for tool calls (not JSON):

markdown
<function=tool_name>
<parameter=param_name>
param_value
</parameter>
</function>

Multiple tool calls are separated by newlines. This format was introduced in Stage-1 and reinforced in Stage-2.

Benchmark Results

Both benchmarks were run on a Modal H200 container loading the base model + adapter with bfloat16 precision.

1. Native Tool-Call Benchmark (44 cases)

The native benchmark evaluates 40 tool-use cases (5 per tool across 8 tools) and 4 no-tool cases.

Table with columns: Metric, Score
Metric	Score
Correct tool selection	79.5%
Required args present	88.6%
Native format valid	100.0%
Correct call count	88.6%
JSON params valid	100.0%
No old JSON format	100.0%
No markdown fences	100.0%
Balanced tags	100.0%

Per-Tool Accuracy

Table with columns: Tool, Accuracy, Args Present, Correct/Total
Tool	Accuracy	Args Present	Correct/Total
`ask_user_question`	100.0%	100.0%	5/5
`edit`	100.0%	100.0%	5/5
`glob`	100.0%	100.0%	5/5

2. BFCL-Compatible Evaluation (10 cases)

The BFCL-style evaluation tests general function-calling with unseen tools (weather, flights, email, stocks, etc.).

Table with columns: Metric, Score
Metric	Score
Correct tool selection	100.0%
Native format valid	100.0%
No old JSON format	100.0%

All 10 BFCL test cases passed with perfect tool selection and format compliance.

Key Findings

Perfect format compliance — 100% native XML format, zero legacy JSON, zero markdown fences, zero extra prose. The Stage-2 adapter fully internalized the native tool-call format.
Strong general tool selection — 6 of 8 tools scored 100% accuracy. ask_user_question, edit, glob, read_file, run_shell_command, and todo_write are fully reliable.
Generalization to unseen tools — The 100% BFCL score confirms the model generalizes its tool-calling ability to functions it has never seen during training.
Weaknesses identified:
- search_code (40%) — the model sometimes substitutes grep_search instead of search_code, or calls with -style arguments.

Usage

python
from peft import PeftModel
from transformers import AutoModelForCausalLM, AutoTokenizer

base_model = "TaimoorSiddiqui/Hopcoder-Mini-9B"
adapter_repo = "TaimoorSiddiqui/Hopcoder-Mini-9B-SWE-Agent-LoRA-H200"

tokenizer = AutoTokenizer.from_pretrained(base_model)
model = AutoModelForCausalLM.from_pretrained(base_model, torch_dtype="auto", device_map="auto")
model = PeftModel.from_pretrained(model, adapter_repo)

Files

Table with columns: File, Description
File	Description
`adapter_config.json`	LoRA configuration
`adapter_model.safetensors`	Adapter weights
`stage2_benchmark_results.json`	Combined benchmark results (native + BFCL)
`stage2_native_benchmark.json`	Full native tool-call benchmark with per-case details
`stage2_bfcl_eval.json`	Full BFCL evaluation with per-case details

Citation

bibtex
@misc{hopcoder-mini-9b-swe-agent-lora-h200,
  author = {TaimoorSiddiqui},
  title = {HopCoder-Mini-9B SWE-Agent LoRA H200},
  year = {2026},
  url = {https://huggingface.co/TaimoorSiddiqui/Hopcoder-Mini-9B-SWE-Agent-LoRA-H200}
}

License

This adapter inherits the license of the base model TaimoorSiddiqui/Hopcoder-Mini-9B.

Hopcoder-Mini-9B-SWE-Agent-LoRA-H200

README

Model Details

Training Data

Native Tool-Call Format

Benchmark Results

1. Native Tool-Call Benchmark (44 cases)

Per-Tool Accuracy

2. BFCL-Compatible Evaluation (10 cases)

Key Findings

Usage

Files

Citation

License

Explore FriendliAI today

README

Model Details

Training Data

Native Tool-Call Format

Benchmark Results

1. Native Tool-Call Benchmark (44 cases)

Per-Tool Accuracy

2. BFCL-Compatible Evaluation (10 cases)

Key Findings

Usage

Files

Citation

License