update endpoint helper files
Browse files- SPECIFICATION.md +134 -0
SPECIFICATION.md
ADDED
|
@@ -0,0 +1,134 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
# Project Specification
|
| 2 |
+
|
| 3 |
+
## 1. Project Name
|
| 4 |
+
|
| 5 |
+
Local Advanced Fine-Tuning Pipeline for Coding LLM
|
| 6 |
+
|
| 7 |
+
## 2. Purpose
|
| 8 |
+
|
| 9 |
+
Provide a fully local, modular workflow to fine-tune a compact coding LLM for:
|
| 10 |
+
- code fixing
|
| 11 |
+
- debugging
|
| 12 |
+
- code explanation
|
| 13 |
+
- response confidence and relevancy signals
|
| 14 |
+
|
| 15 |
+
## 3. Functional Requirements
|
| 16 |
+
|
| 17 |
+
### FR-1 Dataset Generation
|
| 18 |
+
- System must generate a JSON dataset with fields:
|
| 19 |
+
- `instruction`
|
| 20 |
+
- `input`
|
| 21 |
+
- `output`
|
| 22 |
+
- `explanation`
|
| 23 |
+
- `confidence`
|
| 24 |
+
- `relevancy`
|
| 25 |
+
- Dataset size must be constrained to 5000-10000 samples.
|
| 26 |
+
|
| 27 |
+
### FR-2 Model Fine-Tuning
|
| 28 |
+
- System must support LoRA fine-tuning on:
|
| 29 |
+
- `Qwen/Qwen2.5-Coder-0.5B-Instruct` (default)
|
| 30 |
+
- Training inputs must be tokenized and formatted from dataset records.
|
| 31 |
+
- Training output must be stored in a configurable output directory.
|
| 32 |
+
|
| 33 |
+
### FR-3 Pipeline Orchestration
|
| 34 |
+
- System must provide a one-command execution script for:
|
| 35 |
+
- dataset generation
|
| 36 |
+
- training
|
| 37 |
+
- optional uploading
|
| 38 |
+
- Pipeline must support skipping individual stages.
|
| 39 |
+
|
| 40 |
+
### FR-4 Local Inference
|
| 41 |
+
- System must generate outputs from local model folder.
|
| 42 |
+
- Inference module must support:
|
| 43 |
+
- LoRA adapter outputs
|
| 44 |
+
- full model outputs
|
| 45 |
+
- Inference output must be valid JSON containing:
|
| 46 |
+
- `code`
|
| 47 |
+
- `explanation`
|
| 48 |
+
- `confidence`
|
| 49 |
+
- `important_tokens`
|
| 50 |
+
- `relevancy_score`
|
| 51 |
+
- `hallucination`
|
| 52 |
+
- `hallucination_check_reason`
|
| 53 |
+
- `latency_ms`
|
| 54 |
+
|
| 55 |
+
### FR-5 HF Upload
|
| 56 |
+
- System must upload trained model artifacts to a user-specified HF repo.
|
| 57 |
+
- Upload should be optional and independently executable.
|
| 58 |
+
- System must support updating an existing HF model repo by uploading to the same `repo_id`.
|
| 59 |
+
|
| 60 |
+
## 4. Non-Functional Requirements
|
| 61 |
+
|
| 62 |
+
### NFR-1 Reliability
|
| 63 |
+
- Scripts must fail with clear error messages for missing files/directories.
|
| 64 |
+
|
| 65 |
+
### NFR-2 Configurability
|
| 66 |
+
- Hyperparameters and paths must be configurable via CLI.
|
| 67 |
+
- Pipeline defaults should be read from `training_config.json`.
|
| 68 |
+
|
| 69 |
+
### NFR-3 Performance
|
| 70 |
+
- Must support limited-sample smoke run for CPU environments.
|
| 71 |
+
- Tokenization must use deterministic fixed-length padding for stable LoRA training labels.
|
| 72 |
+
- Inference should support deterministic mode by default for stable outputs.
|
| 73 |
+
|
| 74 |
+
### NFR-4 Maintainability
|
| 75 |
+
- Modules must remain decoupled and single-purpose where possible.
|
| 76 |
+
- Documentation must include setup and run commands.
|
| 77 |
+
|
| 78 |
+
## 5. Input/Output Contracts
|
| 79 |
+
|
| 80 |
+
### Dataset Generator
|
| 81 |
+
- Input:
|
| 82 |
+
- `--size` (int, 5000-10000)
|
| 83 |
+
- `--out` (path)
|
| 84 |
+
- Output:
|
| 85 |
+
- JSON training file at `--out`
|
| 86 |
+
|
| 87 |
+
### Trainer
|
| 88 |
+
- Input:
|
| 89 |
+
- dataset file path
|
| 90 |
+
- model name
|
| 91 |
+
- hyperparameters
|
| 92 |
+
- Output:
|
| 93 |
+
- trained model artifacts in `output_dir`
|
| 94 |
+
|
| 95 |
+
### Inference
|
| 96 |
+
- Input:
|
| 97 |
+
- local model path
|
| 98 |
+
- prompt
|
| 99 |
+
- max new tokens
|
| 100 |
+
- Output:
|
| 101 |
+
- structured JSON to stdout
|
| 102 |
+
- Contract:
|
| 103 |
+
- required keys: `code`, `explanation`, `confidence`, `important_tokens`, `relevancy_score`, `hallucination`, `hallucination_check_reason`, `latency_ms`
|
| 104 |
+
|
| 105 |
+
### Upload
|
| 106 |
+
- Input:
|
| 107 |
+
- model directory path
|
| 108 |
+
- HF repo id
|
| 109 |
+
- Output:
|
| 110 |
+
- model artifacts uploaded to HF repo
|
| 111 |
+
|
| 112 |
+
## 6. Default Configuration
|
| 113 |
+
|
| 114 |
+
- Model: `Qwen/Qwen2.5-Coder-0.5B-Instruct`
|
| 115 |
+
- Dataset size: `8000`
|
| 116 |
+
- Epochs: `3`
|
| 117 |
+
- Batch size: `2`
|
| 118 |
+
- Learning rate: `1e-4`
|
| 119 |
+
- Max length: `512`
|
| 120 |
+
|
| 121 |
+
## 7. Validation Criteria
|
| 122 |
+
|
| 123 |
+
Project is considered runnable when:
|
| 124 |
+
- all scripts compile
|
| 125 |
+
- dataset generation succeeds
|
| 126 |
+
- a smoke training run completes
|
| 127 |
+
- inference returns valid JSON payload with required keys
|
| 128 |
+
- upload script accepts valid model dir and repo id
|
| 129 |
+
|
| 130 |
+
## 8. Known Constraints
|
| 131 |
+
|
| 132 |
+
- CPU training is slow for full dataset runs.
|
| 133 |
+
- HF login/token is required for upload.
|
| 134 |
+
- Output quality depends heavily on dataset diversity and quality.
|