tsft-rag-qwen2.5-1.5b-instruct/README.md

---
license: apache-2.0
library_name: transformers
pipeline_tag: text-generation
base_model: Qwen/Qwen2.5-1.5B-Instruct
base_model_relation: finetune
language:
  - de
  - en
tags:
  - rag
  - retrieval-augmented-generation
  - task-specific-fine-tuning
  - full-fine-tuning
  - qwen2.5
  - text-generation
  - structured-output
  - abstention
  - research
  - tsft-rag
  - model-card-v1
---

# TSFT-RAG Qwen2.5-1.5B-Instruct

**Task-Specific Full Fine-Tuning for Retrieval-Augmented Generation**

TSFT-RAG Qwen2.5-1.5B-Instruct is a full-parameter fine-tuned derivative of
[Qwen/Qwen2.5-1.5B-Instruct](https://huggingface.co/Qwen/Qwen2.5-1.5B-Instruct).

Part of the [TSFT-RAG model family](https://huggingface.co/sascha-frank-ai-research) for Retrieval-Augmented Generation across Gemma, Qwen and Llama architectures.

The objective is to investigate how supervised full fine-tuning changes the behaviour of modern language models when used as Retrieval-Augmented Generation (RAG) systems.

---

# TSFT-RAG Model Series

| Model | Parameters | Status |
|--------|-----------:|--------|
| [Qwen2.5-0.5B-Instruct](https://huggingface.co/sascha-frank-ai-research/tsft-rag-qwen2.5-0.5b-instruct)  | 0.5B | Released |
| **Qwen2.5-1.5B-Instruct** | **1.5B** | **Released** |
| [Qwen2.5-7B-Instruct](https://huggingface.co/sascha-frank-ai-research/tsft-rag-qwen2.5-7b-instruct) | 7B | Released |
| [Qwen2.5-14B-Instruct](https://huggingface.co/sascha-frank-ai-research/tsft-rag-qwen2.5-14b-instruct) | 14B | Released |

---

The model was trained to improve several behaviours that are important in practical
RAG systems:

- answering from supplied context;
- abstaining when the context does not support an answer;
- following task-specific output requirements;
- producing structured JSON output;
- generating source-aware and citation-oriented responses.

The main observed gains are in hard-negative handling, structured-output validity,
and citation-related behaviour. Performance does not improve uniformly across all
metrics; in particular, the measured grounded-QA score is lower than that of the
unchanged base model in the reported evaluation.

## Model Details

| Property | Value |
|---|---|
| Model family | TSFT-RAG |
| Base model | Qwen/Qwen2.5-1.5B-Instruct |
| Architecture | Qwen2ForCausalLM |
| Parameters | approximately 1.5B |
| Training method | full-parameter supervised fine-tuning |
| Primary language of the training corpus | German |
| Context length supported by the configuration | 32,768 tokens |
| Weight dtype | bfloat16 |
| Framework | Transformers |
| License | Apache-2.0 |

## Intended Use

This model is intended for research and experimental use in RAG-oriented workflows.

Suitable use cases include:

- context-grounded question answering;
- evaluation of abstention behaviour;
- generation of structured RAG responses;
- source-aware answer generation;
- comparison of base and fully fine-tuned small language models;
- local or resource-constrained RAG experiments.

The model should be used with retrieved evidence or other explicitly supplied context.
It is not intended to serve as a standalone factual knowledge source.

## Training Data

The model was trained on a fixed, task-specific German-language corpus designed for
RAG-oriented instruction following.

The training examples cover:

- answerable context-grounded questions;
- hard-negative and unanswerable questions;
- multiple assistant roles and response styles;
- analysis tasks for main topic, key message, and main arguments;
- JSON and short-answer output formats;
- citation-aware responses.

The complete raw training corpus is not distributed with this checkpoint. The
accompanying repository documents the data-processing, training, and evaluation
pipeline.

## Training Procedure

This checkpoint was produced using full-parameter supervised fine-tuning rather than
LoRA, QLoRA, or another adapter-based method.

| Training argument | Value |
|---|---:|
| Epochs | 3 |
| Learning rate | 1e-6 |
| Per-device train batch size | 1 |
| Per-device evaluation batch size | 1 |
| Gradient accumulation steps | 16 |
| Effective batch size per process | 16 |
| Optimizer | paged AdamW 8-bit |
| Scheduler | cosine |
| Warmup ratio | 0.08 |
| Weight decay | 0.01 |
| Maximum gradient norm | 1.0 |
| Precision | bfloat16 |
| FP16 | disabled |
| Packing | disabled |
| Seed | 42 |

The recorded training runtime was approximately **21,789 seconds** (about **6 h 3 min**),
with a final reported training loss of **1.0738**.

## Evaluation

The fine-tuned model and the unchanged base model were evaluated on the same held-out
test set of **1,940 instances**.

| Metric | Base model | TSFT-RAG | Difference |
|---|---:|---:|---:|
| Aggregate score | 0.3321 | 0.4308 | +0.0987 |
| Hard-negative handling | 0.0056 | 0.3418 | +0.3362 |
| Grounded QA | 0.5289 | 0.4803 | -0.0486 |
| Analysis: main topic | 0.5593 | 0.5518 | -0.0075 |
| Analysis: key message | 0.5210 | 0.4810 | -0.0400 |
| Analysis: main arguments | 0.4549 | 0.3998 | -0.0551 |
| JSON validity | 0.7477 | 0.8862 | +0.1386 |
| Citation precision | 0.0000 | 0.6502 | +0.6502 |
| Citation recall | 0.0000 | 0.6502 | +0.6502 |

The strongest change is in hard-negative handling: the unchanged base model correctly
rejected 4 of the 708 hard-negative cases under the benchmark criterion, while the
fine-tuned model correctly rejected 242. JSON validity and citation-related scores also
increased substantially.

These results should be interpreted as task-specific benchmark outcomes rather than
general-purpose capability claims. The model improves substantially in abstention, JSON
validity, and citation-related behaviour, while grounded-QA and the three analysis-task
scores decrease relative to the unchanged base model.

Further evaluation details, scripts, and analysis are available in the accompanying
repository and paper.

## Usage

```python
import torch
from transformers import AutoModelForCausalLM, AutoTokenizer

model_id = "sascha-frank-ai-research/tsft-rag-qwen2.5-1.5b-instruct"

tokenizer = AutoTokenizer.from_pretrained(model_id)

model = AutoModelForCausalLM.from_pretrained(
    model_id,
    torch_dtype=torch.bfloat16,
    device_map="auto",
)

messages = [
    {
        "role": "system",
        "content": (
            "Du bist ein sachlicher RAG-Assistent. "
            "Beantworte die Frage präzise nur auf Grundlage des Kontexts. "
            "Wenn der Kontext die Antwort nicht belegt, sage dies ausdrücklich."
        ),
    },
    {
        "role": "user",
        "content": (
            "Kontext:\n"
            "Ein Retrieval-Modul liefert relevante Dokumentpassagen an das Sprachmodell.\n\n"
            "Frage:\n"
            "Welche Komponente stellt dem Sprachmodell die Dokumentpassagen bereit?"
        ),
    },
]

text = tokenizer.apply_chat_template(
    messages,
    tokenize=False,
    add_generation_prompt=True,
)

inputs = tokenizer(text, return_tensors="pt").to(model.device)

with torch.inference_mode():
    output = model.generate(
        **inputs,
        max_new_tokens=256,
        do_sample=False,
    )

generated_tokens = output[0, inputs["input_ids"].shape[1]:]

print(
    tokenizer.decode(
        generated_tokens,
        skip_special_tokens=True,
    )
)
```

## Prompting Notes

The model was trained with Qwen's chat template and should be used through
`tokenizer.apply_chat_template(...)`.

For RAG use, prompts should clearly separate:

1. the behavioural instruction;
2. the retrieved context;
3. the user question;
4. any required output format.

For deterministic evaluation, use greedy decoding or another fixed decoding setup.

## Limitations

- The model may produce incorrect, incomplete, or unsupported answers.
- Fine-tuning does not guarantee faithful use of retrieved evidence.
- Retrieval quality remains a major determinant of answer quality.
- The model may fail to abstain even when evidence is insufficient.
- Citation-like output does not guarantee that a cited passage truly supports a claim.
- The 1.5B parameter scale still limits reasoning depth and general robustness.
- Performance outside the evaluated RAG task formats has not been systematically established.
- The training corpus is primarily German, so behaviour may differ in other languages.
- The model inherits limitations and biases from the Qwen2.5 base model and from the
  task-specific training data.
- The observed reductions in grounded-QA and analysis-task scores indicate real
  trade-offs and should not be ignored when selecting the model for deployment.

The model should not be used for high-stakes medical, legal, financial, safety-critical,
or administrative decisions without independent validation and human oversight.

## Ethical Considerations

Retrieved documents may contain inaccurate, biased, confidential, or outdated
information. Such content can be reproduced or amplified by the model.

Users are responsible for:

- validating retrieval sources;
- protecting confidential material;
- testing model behaviour in the intended domain;
- monitoring unsupported or misleading output;
- complying with applicable law, policy, and licensing requirements.

## Relationship to the Research Project

This checkpoint is one model from a controlled cross-family study of task-specific
full fine-tuning for RAG. The study compares base and fine-tuned configurations from
the Gemma, Llama, and Qwen model families across parameter scales from approximately
0.5B to 14B.

Code, training scripts, evaluation scripts, and documentation:

https://github.com/frankmst/rag-task-specific-full-finetuning

## Base Model and License

This model is derived from
[Qwen/Qwen2.5-1.5B-Instruct](https://huggingface.co/Qwen/Qwen2.5-1.5B-Instruct),
which is distributed under the Apache License 2.0.

This derivative checkpoint is also released under the Apache License 2.0. Users
should review the license file included in this repository and the licensing
information of the base model.

## Citation

A formal paper citation will be added after publication of the accompanying paper.

Until then, please cite the software repository:

```bibtex
@software{frank_tsft_rag_2026,
  author       = {Sascha Frank},
  title        = {Task-Specific Full Fine-Tuning for Retrieval-Augmented Generation},
  year         = {2026},
  url          = {https://github.com/frankmst/rag-task-specific-full-finetuning}
}
```

## Project

The complete TSFT-RAG research project, including training scripts,
evaluation pipeline, benchmark methodology and documentation, is available at:

**Project repository**
https://github.com/frankmst/rag-task-specific-full-finetuning

**TSFT-RAG model family**
https://huggingface.co/sascha-frank-ai-research

---

## Author

**Sascha Frank**

Independent AI Researcher

**ORCID**
https://orcid.org/0000-0002-0588-0081

**GitHub**
https://github.com/frankmst

**Hugging Face**
https://huggingface.co/sascha-frank-ai-research

## Links

- Project repository: https://github.com/frankmst/rag-task-specific-full-finetuning
- Base model: https://huggingface.co/Qwen/Qwen2.5-1.5B-Instruct
- Project website: https://sascha-frank.com
- Paper: forthcoming