Update README.md
Browse files
README.md
CHANGED
|
@@ -1,3 +1,95 @@
|
|
| 1 |
---
|
| 2 |
license: mit
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 3 |
---
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
---
|
| 2 |
license: mit
|
| 3 |
+
task_categories:
|
| 4 |
+
- text-generation
|
| 5 |
+
language:
|
| 6 |
+
- en
|
| 7 |
+
pretty_name: Strata-distilled
|
| 8 |
+
size_categories:
|
| 9 |
+
- n<1K
|
| 10 |
+
tags:
|
| 11 |
+
- synthetic
|
| 12 |
+
- distillation
|
| 13 |
+
- reasoning
|
| 14 |
+
- chain-of-thought
|
| 15 |
---
|
| 16 |
+
|
| 17 |
+
# Strata-distilled
|
| 18 |
+
|
| 19 |
+
A small **test** dataset (~50 conversations) of synthetic chat data with reasoning traces at several thinking levels. It can be used to fine-tune smaller thinking models.
|
| 20 |
+
|
| 21 |
+
> **Test release.** This is a proof-of-concept run of a larger generation pipeline, not a finished dataset.
|
| 22 |
+
|
| 23 |
+
## Disclaimer
|
| 24 |
+
|
| 25 |
+
- **Generated by a heavily quantized model.** All data was produced by **`swift-1.5-iq3_xxs`**, an IQ3_XXS (roughly 3-bit) quantization. Heavy quantization lowers answer quality and reasoning reliability, so expect factual errors, flawed logic, weak code and inconsistent formatting.
|
| 26 |
+
- **Not checked before publishing.** The samples are raw model output. They were not reviewed by a human or verified for correctness, safety or bias. Do not treat the answers or the reasoning as ground truth.
|
| 27 |
+
- Use it to test a training pipeline. Filter and validate it before relying on it for anything serious.
|
| 28 |
+
|
| 29 |
+
## How it was made
|
| 30 |
+
|
| 31 |
+
One teacher model (`swift-1.5-iq3_xxs`, served locally through an OpenAI-compatible API) played both sides of every conversation:
|
| 32 |
+
|
| 33 |
+
1. **The user.** The model was prompted to write casual, human-sounding messages. Each conversation got a random persona, writing style and message length, and a concrete topic idea brainstormed by the model, so messages read like real chat input (lowercase, slang, typos, terse or rambling).
|
| 34 |
+
2. **The assistant.** The model answered each turn, with thinking turned on or off depending on the sample's thinking level.
|
| 35 |
+
|
| 36 |
+
Conversations have 1 to 4 turns, and some start with a system prompt.
|
| 37 |
+
|
| 38 |
+
**Topics:** coding, MMLU-style multiple-choice questions across the 57 MMLU subjects, world knowledge, math and reasoning, writing and language, practical advice, instruction following and casual chat.
|
| 39 |
+
|
| 40 |
+
## Thinking levels
|
| 41 |
+
|
| 42 |
+
Each sample has a `thinking_level` field:
|
| 43 |
+
|
| 44 |
+
| Level | Meaning |
|
| 45 |
+
|---|---|
|
| 46 |
+
| `off` | No reasoning. The assistant turn begins with an empty `<think>` block. |
|
| 47 |
+
| `low` | Very short reasoning. |
|
| 48 |
+
| `medium` | Focused, step-by-step reasoning. |
|
| 49 |
+
| `high` | Long, thorough reasoning with double-checking. |
|
| 50 |
+
|
| 51 |
+
Levels were requested through the prompt, and samples whose reasoning length didn't fit the requested level were discarded. They are approximate labels, not guarantees.
|
| 52 |
+
|
| 53 |
+
## Format
|
| 54 |
+
|
| 55 |
+
The file is a JSON array. Each element looks like this:
|
| 56 |
+
|
| 57 |
+
```json
|
| 58 |
+
{
|
| 59 |
+
"id": "a1b2c3d4e5f6",
|
| 60 |
+
"category": "coding",
|
| 61 |
+
"subcategory": "Python",
|
| 62 |
+
"thinking_level": "medium",
|
| 63 |
+
"num_turns": 2,
|
| 64 |
+
"messages": [
|
| 65 |
+
{"role": "system", "content": "You are a helpful assistant."},
|
| 66 |
+
{"role": "user", "content": "..."},
|
| 67 |
+
{"role": "assistant", "content": "<think>\n...reasoning...\n</think>\n\nfinal answer"}
|
| 68 |
+
]
|
| 69 |
+
}
|
| 70 |
+
```
|
| 71 |
+
|
| 72 |
+
- `system` messages are optional.
|
| 73 |
+
- Assistant turns follow the Qwen3 chat format. The reasoning is inside `<think>...</think>`, followed by the answer. With `off`, the block is empty (`<think>\n\n</think>\n\n`).
|
| 74 |
+
- `messages` is the standard chat format used by most fine-tuning libraries (TRL, Unsloth, Axolotl and others).
|
| 75 |
+
|
| 76 |
+
## Usage
|
| 77 |
+
|
| 78 |
+
```python
|
| 79 |
+
from datasets import load_dataset
|
| 80 |
+
|
| 81 |
+
ds = load_dataset("json", data_files="swift_distill_dataset.json", split="train")
|
| 82 |
+
print(ds[0]["messages"])
|
| 83 |
+
```
|
| 84 |
+
|
| 85 |
+
## Limitations
|
| 86 |
+
|
| 87 |
+
- Tiny sample size, so it is not suitable for real fine-tuning on its own.
|
| 88 |
+
- Single teacher model, so its mistakes and style biases are copied into the data.
|
| 89 |
+
- Reasoning traces come from a heavily quantized model and may be wrong even when the final answer is right (or the reverse).
|
| 90 |
+
- English only.
|
| 91 |
+
- Simulated users are imitations of human messages, not real user data.
|
| 92 |
+
|
| 93 |
+
## License
|
| 94 |
+
|
| 95 |
+
MIT, as declared in the metadata above.
|