| # sft-forge-ensemble (v0.3) |
|
|
| Single pareto-pruned SFT dataset ensembled from multiple high-quality sources. |
| Generated by **sft-forge** — t. |
|
|
| ## Pipeline |
| ingest -> tag every row -> programmatic refine -> GLOBAL dedup (one minhash |
| index across all sources) -> per-slice pareto prune (quality vs REAL gpt2 |
| token cost) -> this file. |
|
|
| ## Sources (rows ingested) |
| | source | rows | |
| |---|---| |
| | a | 1 | |
|
|
| ## Survival after dedup + pareto (kept/ingested) |
| | source | kept | |
| |---|---| |
| | a | 1/1 | |
|
|
| - Kept rows: **1** (0 dominated rows pruned) |
| - Real token total (gpt2 BPE): **123** |
| - Slices: 1 capability slices |
|
|
| ## Token budget by source (gpt2, kept rows only) |
| | source | tokens | share | |
| |---|---|---| |
|
|
|
|
| Deep agentic trajectories dominate raw token share by design (they are the |
| rows small models cannot synthesize); rebalance a training mix by dropping |
| or down-weighting `src:pyromind_long`-tagged rows via the tag filter. |
|
|
| ## Format |
| `train.jsonl`: {"id", "tags", "messages" (assistant turns loss-mask-ready via |
| `trainable`), "approx_tokens", "gpt2_tokens", "meta"}. Filter by tags, never |
| re-split. `gpt2_tokens` is the true BPE count — trust it for packing. |
|
|
