---
tags:
- whittle
- research
- training-checkpoints
---

# whittle-dev

> ### ☕ Support this work
> Whittle is built by one person on a grocery budget and rented GPU hours. If this research is useful to you, or you want to see it finished:
> **[ko-fi.com/davida81328](https://ko-fi.com/davida81328)**. Every hour of GPU time goes straight into the next checkpoint, and every checkpoint, table and log lands in these repos.

Nightly training checkpoints (trainer state: HC/PLE/LoRA/gates) for Whittle-Next runs on Colab.
`colab1/latest.pt` is overwritten as the run progresses; `colab1/step<N>.pt` are milestones (the earlier
note on this card said every 1000 steps; the tree holds `step200.pt` to `step4400.pt`, every 200 steps, as
the Whittle-Next-27B-A3B card states). Load with the
trainer (`RESUME_CKPT=`), not as a standalone model.

## Status

Development artefacts, not a release. These are the run directories referenced by the
[Whittle-Next-27B-A3B](https://huggingface.co/logic65/Whittle-Next-27B-A3B) and
[Whittle-Qwen-3.8-35B-A3B](https://huggingface.co/logic65/Whittle-Qwen-3.8-35B-A3B) cards: `colab1/` is the
v4.4 run of Whittle-Next-27B-A3B (its checkpoints and run-end table), and `tbl1/`, `lw2/`, `lw5/` and
`agentfix2/` are the checkpoints of the same names described on the Whittle-Qwen-3.8-35B-A3B card. The
other directories (`colab2/`, `lw1/`, `lw3/`, `lw4/`, `agentfix/`, `ai2-archive/`) are runs of the same
line that no card describes.

## What is here

- `colab1/` — `step<N>.pt` (22 milestones), `latest.pt`, `ngram_table_final.npy`.
- `colab2/` — `step<N>.pt`, `latest.pt`, `ngram_table_final.npy`, `eval/` (CE curve, layer-wise eval logs and JSON, export log, `TRAIN.log`).
- `lw1/` — `step<N>.pt`, `latest.pt`, `ngram_table_final.npy`, `eval/` (CE curve, layer-wise eval logs and JSON, export log, `TRAIN.log`).
- `lw2/`, `lw3/`, `lw4/`, `lw5/` — `step<N>.pt`, `latest.pt`, `ngram_table.npy`, `ple_hash.json`, `TRAIN.log`, `eval/` (CE curve, layer-wise eval logs and JSON, `math60_replies.jsonl` + `math60_score.txt`, export/convert logs) and one Q8_0 GGUF per run: `Whittle-Qwen-3.8-35B-A3B-lw2-Q8_0.gguf`, `…-lw3-…`, `…-lw4-…`, `…-lw5-…`.
- `tbl1/` — `step<N>.pt`, `latest.pt`, `ngram_table.npy`, `ple_hash.json`, `eval/` (GGUF build and quantisation logs, `gguf_sizes.json`).
- `agentfix/` — `latest.pt`, `step200.pt`.
- `agentfix2/` — `latest.pt`, `step300.pt`, `bf16/` (full weights, 14 safetensors shards + config, tokenizer and chat template) and `Whittle-Qwen-3.8-35B-A3B-agentfix2-Q8_0.gguf`.
- `ai2-archive/` — archived trainer states: `reader7/`, `reader9/`, `reader11/`, `reader12/` (`TRAIN.log`, `ngram_table.npy`, `ple_hash.json`, and `keep_final.pt` for all but reader9), `v3ref/` (`resume_from.pt`, `ple_hash.json`), `v4/` (`colab_onpolicy_next_step42.pt`).

The checkpoints are trainer state, not standalone models: the trainer, exporter and the frozen body they
apply to are described on the two model cards above.