| --- |
| license: apache-2.0 |
| tags: |
| - code |
| - tool-use |
| - reasoning |
| - preference-data |
| - expert-routing |
| - reproducibility |
| - autonoma |
| --- |
| |
| # Autonoma Fable router curriculum — reproducibility package |
|
|
| This public repository documents and reproduces the router-only curriculum for an experimental LFM2.5-2.6B Fable hybrid. It intentionally does **not** redistribute records derived from a source whose license is unknown. |
|
|
| ## Included |
|
|
| - `fable-router-curriculum.v2.json`: immutable training/data policy. |
| - `build_fable_router_curriculum.py`: deterministic local builder. |
| - `REPRODUCIBILITY_MANIFEST.json`: pinned sources, input/output hashes, row counts, exclusions, and build command. |
| - `training/`: the non-routing router shell, pure provenance validators, and free-GPU structural-smoke runner. |
| - `fable-router-cloud-smoke.v1.json`: immutable host, bank, curriculum, hash, hardware, and evidence gates for that runner. |
|
|
| Given independently authorized access to the two public sources, these files reproduce the exact materialized artifacts by SHA-256. They also expose the split, deduplication, overflow, anti-loop, duplicate-tool, benefit-gated routing, and anti-collapse decisions for review. |
|
|
| ## Materialized result |
|
|
| The internally verified build produced: |
|
|
| | File | Rows | SHA-256 | |
| |---|---:|---| |
| | `sft-train.jsonl` | 6,084 | `10a72911887aa642339b78c11ed12f3540ada2076a81dc13fc99566bf522bd13` | |
| | `sft-validation.jsonl` | 341 | `b2925d1699a83ea48649905a6e922e311f34e355bdbf22b0b573692358643e66` | |
| | `sft-test.jsonl` | 330 | `7ddd0e5865ba54c12d1aca9ca8b87fa078742d2a115ab56137125757ac343421` | |
| | `preference-train.jsonl` | 4,099 | `9924541e4a5b40710e14a370119d7571e71bb2f881433b6bd5807cc1c76dc8a4` | |
|
|
| The split unit is normalized prompt hash for reasoning data and immutable source trajectory SHA-256 for agent/coding data. No identity crosses a split. Frozen evaluation prompts and answers are excluded from the builder. |
|
|
| ## Source policy |
|
|
| - [`greghavens/fable-5-coding-and-debugging-traces`](https://huggingface.co/datasets/greghavens/fable-5-coding-and-debugging-traces) at `c63e82adec30798edcbd6e1dcb0014d2b15de236` declares CC-BY-4.0. |
| - [`HelioAI/Claude-Fable-5-5500x`](https://huggingface.co/datasets/HelioAI/Claude-Fable-5-5500x) at `440267fbdb1b00a40216e7233dbce25530a0ed09` declares `license: unknown` and explicitly tells users to verify compliance before redistribution or derivative training. |
|
|
| Public availability is not itself a redistribution license. Accordingly, this public package contains references and reproducibility code, not copied Fable reasoning records. The materialized artifacts remain in a private companion repository pending a compatible license or a rebuild without that source. |
|
|
| ## Reproduction |
|
|
| Place the pinned source files at the paths declared in the builder, provide the pinned Fable tokenizer, then run: |
|
|
| ```bash |
| python scripts/bench/build_fable_router_curriculum.py |
| ``` |
|
|
| The builder rejects malformed or overlong examples, deduplicates reasoning prompts, samples cumulative agent prefixes by immutable trajectory, creates train-only anti-loop preference pairs, and writes a provenance-bound `result.json`. Final training must retokenize with the actual chat template and reject overflow above 2,048 tokens rather than truncate it. |
|
|
| Publication of this package does not authorize training, routing, deployment, or promotion in Autonoma. |
|
|
| ## Structural smoke boundary |
|
|
| The structural smoke validates every one of a bank's 2,880 tensor records, then materializes only layer 29's 32 experts around the real frozen Fable host. It checks exact host-disabled identity, strict trainable-parameter isolation, nonzero router/scale gradients, explicit host-only and positive-benefit counterfactual targets, router checkpoint parity, and persisted resource telemetry. This is intentionally small enough for a free 16 GiB Colab/Kaggle GPU. |
|
|
| Passing it proves that a donor bank is mechanically trainable. It does **not** prove full-30-layer memory fit, final quality, throughput, or promotion fitness, and it cannot authorize a training campaign. Every run is uploaded to a private evidence dataset so an ephemeral notebook cannot turn an incomplete or failed result into a pass. |
|
|
