File size: 5,880 Bytes
daf6d21 832ea2d 06427c3 daf6d21 832ea2d daf6d21 2d0f692 e8acada daf6d21 36d0ac0 06427c3 36d0ac0 daf6d21 2d0f692 daf6d21 06427c3 daf6d21 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 | ---
license: apache-2.0
library_name: transformers
tags:
- text-generation
- sequential-fine-tuning
- lora
- multi-dataset-model
pipeline_tag: text-generation
widget:
- text: "Once upon a time, in a land far away,"
---
# Sequentially Fine-Tuned Language Model: jnjj/xd_v1
## Model Description
This repository hosts a language model that is being **sequentially fine-tuned** using Low-Rank Adaptation (LoRA) on a diverse range of datasets from the Hugging Face Hub.
The process starts from the base model `jnjj/multi-dataset-model` (or its last fine-tuned state from this repository) and continuously adapts by merging LoRA weights after each dataset training cycle.
This experiment aims to build a model with broad, cumulatively acquired knowledge.
**Current Base for Fine-Tuning:** [jnjj/multi-dataset-model](https://huggingface.co/jnjj/multi-dataset-model)
The fully merged model weights and tokenizer are updated periodically at the root of this repository.
## Training Methodology
- **Iterative Fine-Tuning:** The model undergoes cycles of training on different dataset configurations.
- **LoRA Integration:** PEFT's LoRA is employed for parameter-efficient fine-tuning. Adapters are merged post-training.
- **Dynamic Dataset Source:** The script iterates through a wide array of Hugging Face Hub datasets.
- **Rapid Iteration Strategy:** Training per dataset configuration is brief (`max_steps=1`), prioritizing breadth of exposure over depth on any single dataset.
## Training Progress
- **Datasets Processed (Successfully trained on at least one config):** 1
- **Text Examples Streamed (Total):** 6
- **Tokens Processed (Total):** 3072
- **Last Successful Model Update:** 2025-05-08 18:02:08 UTC
### Evaluation Snapshot (Approximate)
- **Current Perplexity (wikitext Subset):** 282.70
- **Perplexity Change:** `-0.51` ⬇️ (vs previous cycle's perplexity)
#### Generated Examples (Qualitative Assessment)
| Category | Input Prompt Snippet | Generated Output Snippet |
|----------------------------|----------------------------------------|----------------------------------------------|
| Story Continuation | `Once upon a time, in a small villag...` | `How do I get the best picture of what we... ` |
| Simple Instruction | `Explain in one sentence why trees a...` | `I have been trying to make progress and ... ` |
| Creative Prompt | `Describe a friendly robot that love...` | `We are pleased to announce the launch of... ` |
| Question Answering (Basic) | `What is the main color of a ripe ba...` | `As an example we've been using the same ... ` |
| Code Generation (Simple Python) | `Write a Python function that takes ...` | `We are looking forward to seeing us in t... ` |
| Reasoning (Simple) | `If a train leaves station A at 10:0...` | `The time of day we were trying to get ou... ` |
#### Standard Benchmarks (via `lighteval`)
_Note: Running standard benchmarks requires a dedicated setup using the `lighteval` harness. The table below shows scores if available in `evaluation_stats.json`, otherwise `N/A`._
**Common Benchmarks**
| Category | Benchmark | # Shots | Metric | This Model (`xd_v1`) | Llama 3.1 70B (Ref) |
|-----------------------|-------------------|---------|----------------|----------------------|---------------------|
| Reasoning & Knowledge | MMLU (Avg) | 5 | acc_norm | `N/A` | 79.3 |
| Reasoning & Knowledge | MMLU-Pro | 5 | acc | `N/A` | 53.8 |
| Reasoning & Knowledge | MATH | 4 | acc | `N/A` | 41.6 |
| Reasoning & Knowledge | TruthfulQA (MC2) | 0 | mc2 | `N/A` | - |
| Reasoning & Knowledge | GPQA Diamond | 0 | acc | `N/A` | 50.5 |
| Code | MBPP | 3 | pass@1 | `N/A` | 66.4 |
| Code | LiveCodeBench | 0 | pass@1 | `N/A` | 33.3 |
| Multilingual | TydiQA | 1 | f1 | `N/A` | 29.9 |
| Multilingual | MGSM | 0 | acc | `N/A` | 91.1 |
## How to Use
Load the model and tokenizer via `transformers`:
```python
from transformers import AutoModelForCausalLM, AutoTokenizer
model_id = "jnjj/xd_v1"
# For local usage after downloading:
# model_id = "./model_files"
tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(model_id)
# model.to("cuda") # if GPU is available
prompt = "Explain the concept of photosynthesis in simple terms:"
inputs = tokenizer(prompt, return_tensors="pt") # .to("cuda" if GPU available)
output_sequences = model.generate(
**inputs,
max_new_tokens=100,
do_sample=True,
temperature=0.7,
top_p=0.9,
pad_token_id=tokenizer.eos_token_id
)
generated_text = tokenizer.decode(output_sequences[0], skip_special_tokens=True)
print(generated_text)
```
## Limitations & Considerations
- This model is an experimental artifact of continuous learning; quality and coherence may vary.
- Biases present in the underlying datasets may be reflected or amplified.
- Performance on specific tasks is not guaranteed and may fluctuate as new datasets are incorporated.
- Intended for research and exploration of sequential fine-tuning dynamics. For rigorous benchmarking, consider using tools like [`lighteval`](https://github.com/huggingface/lighteval).
## Disclaimer
This model is provided as-is. It may generate inaccurate, biased, or otherwise problematic content. Users should exercise discretion.
|