Datasets:
File size: 3,353 Bytes
34a4f90 | 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 | ---
license: apache-2.0
task_categories:
- text-generation
language:
- en
tags:
- distillation
- reasoning-traces
- function-calling
- synthetic
size_categories:
- 10K<n<100K
---
# scram-curriculum: the training set behind Scram-0.8B
17,837 verified teacher traces. This is the *complete* supervised fine-tuning set used
to train [negativevoid/Scram-0.8B-6bit](https://huggingface.co/negativevoid/Scram-0.8B-6bit),
published so the recipe is checkable rather than merely described.
## How it was made
Traces were generated by Qwen3.5-4B and Qwen3.5-9B, then **kept only if the final
answer exactly matched gold**. Roughly 7% of generations were discarded by that filter.
Verification is final-answer-only: no one checked that the intermediate reasoning is
sound, so a trace can reach the right answer by a wrong route. Treat the reasoning as
teacher output, not as ground truth.
## Composition
| kind | rows | what it teaches |
|---|---|---|
| reason | 10,612 | multi-step chains (arithmetic 1,241 · syllogism 1,653 · ordering 1,566 · boolean 1,944 · codetrace 1,343 · gsm8k 3,665) |
| tool | 1,595 | JSON-convention tool calls |
| copy | 1,200 | retrieval-grounded answering from notes |
| pytool | 1,000 | pythonic call convention |
| negation | 800 | scope and negation handling |
| distract | 600 | irrelevant-note resistance |
| synth | 600 | multi-note synthesis |
| factory-pos / factory-ref | 400 / 300 | schema-driven call construction and its refusal corner |
| missing-param | 400 | asking for absent arguments instead of inventing them |
| refusal | 330 | declining out-of-scope requests |
The refusal splits are deliberately class-separated. We found the classes do **not**
cross-transfer: teaching wrong-tool refusal does nothing for missing-param refusal.
Class coverage mattered more than data volume.
## Provenance and licensing
- **3,665 rows (21%) are GSM8K-derived** — questions from
[openai/gsm8k](https://huggingface.co/datasets/openai/gsm8k) *train* split (MIT),
with traces generated by us. **Consequence: any GSM8K evaluation of the resulting
model measures trained capability, not transfer.** Disjointness from the frozen
probe is asserted by exact question match at build time.
- The remaining reasoning items are **fully synthetic with zero real-world entities**
(invented nouns like "Kreskarnpre", "zlupsnus"), so they cannot be answered from
memorized world knowledge.
- Tool schemas are synthetic. The model consequently underperforms on real-world
function inventories (BFCL simple AST 46 vs Hammer2.1-0.5b's 66) — if you fine-tune
on this, expect to add your own domain schemas.
- Teacher outputs come from Qwen3.5 (Apache-2.0). This dataset is Apache-2.0.
## Eval hygiene
Checked at publication: **zero overlap** between these 17,809 unique prompts and every
held-out suite used to evaluate the model (reason n=300, heldout n=80, gsm8k probe
n=100, know n=301). The evaluation instruments themselves are deliberately **not**
published — releasing them would contaminate them for everyone.
## Format
One JSON object per line:
```json
{"question": "...", "trace": "...", "answer": "D", "family": "ordering", "kind": "reason"}
```
Rows of `kind: tool`/`pytool`/`refusal` etc. use `prompt` instead of `question` and
carry no `answer` field (the trace is the target).
|