dim014/deepseek-r1-finetuned

Model Card for Model ID

Model Details

Short Description

The model is a fine-tuned version of the DeepSeek-R1 model, specifically adapted for generating Russian user manuals and instructions for UI forms. It is instruction-tuned and uses LoRA (Low-Rank Adaptation) for efficient parameter-efficient fine-tuning. The model takes UI form descriptions as input and outputs step-by-step instructions in Russian, making it a valuable tool for creating clear and concise user documentation for software interfaces.

Model Description

• Shared by: dim014
• Model type: Causal language model (instruction-tuned, LoRA adapters)
• Language(s) (NLP): Russian (primary), limited English support
• Finetuned from model: deepseek-ai/DeepSeek-R1-Distill-Qwen-1.5B

Model Sources

• Repository: deepseek-r1-finetuned
• Demo: TBC

Uses

Direct Use

• Generate Russian user manuals and step-by-step instructions for UI forms (fields, buttons, dialogs).
• Assist technical writers in drafting consistent, clear end-user documentation in Russian.
• Produce help texts, onboarding content, and tooltips for web/desktop interfaces.

Downstream Use

• Integrate into documentation generation pipelines or help centers.
• Further fine-tune on domain-specific UIs (e.g., banking, healthcare, enterprise).
• Embed into chatbots that explain UI behavior and form usage.

Out-of-Scope Use

• Safety-critical decision-making (medical, legal, financial).
• General-purpose creative writing or code generation.
• High-quality multilingual generation beyond Russian without additional fine-tuning.
• Any usage that conflicts with the base model or dataset licenses.

Bias, Risks, and Limitations

• May reflect biases from the base model and training data.
• Can produce verbose or repetitive instructions for simple forms.
• May hallucinate steps or UI elements not present in the input.
• Quality may degrade on highly specialized domains or non-Russian inputs.

Recommendations

Users (both direct and downstream) should be made aware of the risks, biases and limitations of the model. More information needed for further recommendations.

• Always perform human review before publishing documentation.
• Prefer Russian inputs; apply post-processing for clarity and conciseness.
• For production, implement guardrails, validation, and prompt templates tailored to your UI.

Training Details

Training Data

• Dataset: dim014/ui-form-user-manual-generation-dataset-rus
• Content: Alpaca-style pairs mapping UI form descriptions to Russian user manuals.
• Size observed in run logs: approximately 1,286 training samples.
• Tokenization: padding/truncation to a maximum of 1,024 tokens.

Training Procedure

Preprocessing

• Alpaca-style prompt template with sections:
  • “Instruction:” + instruction
  • Optional “Input:” + input
  • “Response:” + target output + EOS
• Tokenization with the base model tokenizer; labels set equal to input_ids for causal LM.

Training Hyperparameters

• Training regime: LoRA (PEFT) fine-tuning on top of deepseek-ai/DeepSeek-R1-Distill-Qwen-1.5B

• LoRA configuration:

  • r: 16
  • lora_alpha: 32
  • lora_dropout: 0.05
  • target_modules: q_proj, k_proj, v_proj, o_proj, gate_proj, up_proj, down_proj
  • bias: none

• Epochs: 3

• Per-device batch size: 2

• Gradient accumulation steps: 8 (effective batch size 16)

• Learning rate: 2e-4

• Scheduler: cosine with warmup ratio 0.1

• Precision: fp16

• Optimizer: AdamW (torch)

• Max sequence length: 1,024

• Gradient checkpointing: enabled

• Logging: W&B; checkpoints every 50 steps (keep last 3)

Speeds, Sizes, Times

• Final train loss: ~0.559
• Train runtime: 3,256.93 seconds (54.3 minutes)
• Train steps per second: ~0.075
• Train samples per second: ~1.185
• Total training samples: 1,286
• GPU memory (example from logs): ~3.63 GB allocated, max allocated ~7.80 GB
• Hardware model not captured in logs; trained in a single-GPU CUDA environment (e.g., Google Colab)

Evaluation

Testing Data, Factors & Metrics

Testing Data

• Qualitative testing on 5 curated prompts representing different UI forms (registration, participation details, package selection, payment, completion).

Factors

• Russian-only prompts.
• Varying form complexity (single-step to multi-step flows).
• Domain: general software UIs.

Metrics

• Automatic metrics (BLEU/ROUGE) were not computed in the provided run logs.
• Qualitative manual checks for fluency, clarity, completeness, and hallucinations.

Results

• The model produced coherent, step-by-step Russian instructions across the 5 example prompts.
• Occasional minor terminology drift or verbosity observed; human editing recommended.

Technical Specifications

Model Architecture and Objective

• Architecture: decoder-only causal transformer (Qwen-style, distilled in DeepSeek-R1).
• Objective: instruction tuning for Russian UI form user manual generation.

Compute Infrastructure

• Single-node, single GPU (CUDA), automatic device placement (device_map="auto").

Hardware

• GPU name not captured in logs; runtime memory usage indicates a consumer/pro cloud GPU.
• Example memory: ~3.63 GB allocated, up to ~7.80 GB max allocated.

Software

• Libraries: transformers, datasets, peft, torch, accelerate, bitsandbytes, wandb
• Tokenizer and model weights from the Hugging Face Hub.

Glossary

• LoRA (Low-Rank Adaptation): Parameter-efficient fine-tuning adding low-rank adapters to selected layers.
• Causal LM: Autoregressive language model generating the next token from previous context.
• Instruction tuning: Fine-tuning to follow natural-language instructions.

Model Card Authors

dim014

Model Card Contact

dim014