33 lines
1.4 KiB
Markdown
33 lines
1.4 KiB
Markdown
|
---
|
||
|
|
license: llama3.2
|
||
|
|
base_model: meta-llama/Llama-3.2-1B-Instruct
|
||
|
|
library_name: transformers
|
||
|
|
tags: [le-harnais, ablation, retrain-only]
|
||
|
|
---
|
||
|
|
|
||
|
|
# le-harnais / ft-medium-jepa-drop75
|
||
|
|
|
||
|
|
> ⚠️ **retrain-only** — this is an ablation checkpoint, **not useful for inference**. It ships only to reproduce / continue the training study. For real use see the hero models: `le-harnais-ft-agentworld-{1b,3b,8b}`, `le-harnais-ft-counsel`.
|
||
|
|
|
||
|
|
Spider NL→SQL JEPA ablation — shows JEPA does NOT help on loose view-alignment.
|
||
|
|
|
||
|
|
- **Base model:** [`meta-llama/Llama-3.2-1B-Instruct`](https://huggingface.co/meta-llama/Llama-3.2-1B-Instruct) — **Built with Llama**; Llama Community License applies.
|
||
|
|
- **Class:** `ablation`
|
||
|
|
- **Training data:** spider (NL→SQL)
|
||
|
|
- **Headline:** JEPA ≤0 on NL→SQL (the negative result; exec-match ~0.287)
|
||
|
|
|
||
|
|
## Reproduce
|
||
|
|
|
||
|
|
```sh
|
||
|
|
cd refs/llm-jepa && .venv/bin/python ../../tools/run_jepa_sweep.py --seeds 82 --eval-cap 300
|
||
|
|
```
|
||
|
|
|
||
|
|
Full recipe, datasets, and eval commands: see `docs/REPRODUCE.md` in the
|
||
|
|
[le-harnais distribution]. Provenance & license: `docs/PROVENANCE.md`.
|
||
|
|
|
||
|
|
## Formats in this repo
|
||
|
|
- `*.safetensors` — bf16 inference weights (serve with `transformers` or le-harnais `lh-serve`/candle).
|
||
|
|
- `*.Q4_K_M.gguf` — portable 4-bit quant (run via `ollama` / `llama.cpp`; Mac-friendly).
|
||
|
|
|
||
|
|
> Orchestration amplifies a capable generator; it does not create competence.
|