typescript-slm-7b-reasoning-full

TypeScript-SLM-7B-Reasoning-Full

TypeScript-SLM-7B-Reasoning is a 7B-parameter DeepSeek-based model fine-tuned for step-by-step TypeScript reasoning. It merges the base model with LoRA adapters and includes GGUF quantization for local/Ollama workflows.

This repository hosts the full merged model plus GGUF (q4_k_m) for lightweight inference.

Model Description

• Base Model: deepseek-ai/DeepSeek-R1-Distill-Qwen-7B
• Model Type: Causal LM (code reasoning)
• Parameters: 7B
• Context Length: Inherits base DeepSeek-R1-Distill-Qwen-7B window
• Fine-tuning: LoRA on TypeScript reasoning/debugging tasks
• License: MIT
• Language: English, TypeScript/JavaScript code
• System Prompt: Focus on step-by-step debugging, refactoring, and design-level explanations before giving the final typed solution.

What it is good at

• ✅ Explaining TypeScript bugs and fixes
• ✅ Refactoring and API design discussions
• ✅ Generating strongly-typed code for React/Next.js/Angular/Node.js
• ✅ Producing clear reasoning traces before final answers

Intended Uses

Primary: TypeScript reasoning, debugging, refactoring, and guided code generation.
Out-of-scope: Arbitrary natural-language chat unrelated to code; safety-sensitive or factual tasks outside TypeScript.

Prompt Examples

"Debug this TypeScript function and explain the bug step by step:\n\nfunction add(a?: number, b?: number) { return a + b; }"

"Design a typed API surface for a Next.js todo service. Explain design choices, then show the final code."

How to Use

Ollama (recommended for local)

ollama create typescript-slm-7b-reasoning -f gguf/Modelfile-q4_k_m
ollama run typescript-slm-7b-reasoning "Explain why this React hook re-renders too often..."

Transformers

from transformers import AutoModelForCausalLM, AutoTokenizer
import torch

model = AutoModelForCausalLM.from_pretrained(
    "sylvester-francis/typescript-slm-7b-reasoning-full",
    torch_dtype=torch.float16,
    device_map="auto",
)
tokenizer = AutoTokenizer.from_pretrained("sylvester-francis/typescript-slm-7b-reasoning-full")

prompt = "Refactor this TypeScript service for better typing and error handling..."
inputs = tokenizer(prompt, return_tensors="pt").to(model.device)
outputs = model.generate(
    **inputs,
    max_new_tokens=512,
    temperature=0.3,
    top_p=0.95,
    do_sample=True,
    pad_token_id=tokenizer.eos_token_id,
)
print(tokenizer.decode(outputs[0], skip_special_tokens=True))

GGUF (llama.cpp)

huggingface-cli download sylvester-francis/typescript-slm-7b-reasoning-full \
  gguf/typescript-slm-7b-reasoning-q4_k_m.gguf --local-dir ./models

./llama-cli -m ./models/gguf/typescript-slm-7b-reasoning-q4_k_m.gguf \
  -p "Explain and fix this TypeScript type error..."

Model Files

• gguf/typescript-slm-7b-reasoning-q4_k_m.gguf (≈4.7GB)
• gguf/Modelfile-q4_k_m (Ollama import)

Training Data (summary)

• Curated TypeScript code from popular GitHub repos (React, Next.js, Angular, Node.js)
• StackOverflow Q&A focused on debugging and reasoning
• Filters for strong typing, framework best practices, and reasoning-rich examples

Training Configuration (LoRA)

Base Model: deepseek-ai/DeepSeek-R1-Distill-Qwen-7B
Method: LoRA fine-tuning
Target Domains: TypeScript reasoning, debugging, refactoring
LoRA Rank / Alpha: tuned for stability and reasoning depth
Optimizer: AdamW
Max Sequence Length: inherits base model context window

Evaluation

Qualitative checks on TypeScript debugging/refactoring prompts show:

• Clear reasoning steps before final code
• Strong type usage and framework-aware patterns
• Concise, actionable fixes

Safety & Limitations

• May generate incorrect code or hallucinate APIs—review before production use.
• Not a security scanner; do not rely on it for vulnerability assessments.
• Avoid non-code or high-stakes factual tasks.

License

MIT for the fine-tuned model; base model license and dataset terms also apply.

Contact

• Maintainer: Sylvester Francis (@sylvester-francis on Hugging Face)
• Issues/feedback: open a discussion on the model repo