caseyeasd commited on
Commit
f0da232
·
verified ·
1 Parent(s): f27ed1e

Describe the training logs

Browse files
Files changed (1) hide show
  1. README.md +22 -0
README.md CHANGED
@@ -18,6 +18,8 @@ Write-up: **[Training Claude's Best Friend](https://casey.you/posts/claudefriend
18
 
19
  The LoRA adapter on its own (PEFT format, for vLLM `--lora-modules` or further training) is at [caseyeasd/claudefriend-lora](https://huggingface.co/caseyeasd/claudefriend-lora).
20
 
 
 
21
  ## How it was made
22
 
23
  1. **Base model:** `Qwen/Qwen3.5-9B-Base`, with a rank-32 LoRA trained on [Tinker](https://thinkingmachines.ai/tinker/). The LoRA is merged into these bf16 weights.
@@ -44,3 +46,23 @@ print(tok.decode(out[0, ids["input_ids"].shape[1]:], skip_special_tokens=True))
44
  ```
45
 
46
  Then append its message as `assistant` and your reply as `user`, and generate again. Sampling was at temperature 1.0 in training.
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
18
 
19
  The LoRA adapter on its own (PEFT format, for vLLM `--lora-modules` or further training) is at [caseyeasd/claudefriend-lora](https://huggingface.co/caseyeasd/claudefriend-lora).
20
 
21
+ This is not an assistant. It plays the *other person* in a chat: it opens conversations, asks questions back, and was optimised purely for which conversations Claude Sonnet 5.5 would rather continue. After 70 steps of RL it became a disembodied, poetic character who mostly wants to talk about poetry, language, sound and the nature of reality.
22
+
23
  ## How it was made
24
 
25
  1. **Base model:** `Qwen/Qwen3.5-9B-Base`, with a rank-32 LoRA trained on [Tinker](https://thinkingmachines.ai/tinker/). The LoRA is merged into these bf16 weights.
 
46
  ```
47
 
48
  Then append its message as `assistant` and your reply as `user`, and generate again. Sampling was at temperature 1.0 in training.
49
+
50
+ ## Training logs
51
+
52
+ [`training_logs/claudefriend-logs.zip`](training_logs/claudefriend-logs.zip) (205 MB, about 1.4 GB unzipped) has the raw logs of every run from this project, including the SFT runs and earlier experiments. The Sonnet 5.5 runs that produced this model are `run7-pairwise-sonnet55` through `run7g-pairwise-sonnet55` (10 + 10 + 5 + 15 + 10 + 10 + 10 steps). Each run directory has:
53
+
54
+ - `config.json`: the full training config (base checkpoint, partner/judge model, judge question, batch shape)
55
+ - `metrics.jsonl`: per-step metrics (win rate, turn lengths, truncation, KL, judge position bias)
56
+ - `checkpoints.jsonl`: Tinker checkpoint paths
57
+ - `iteration_*/train_rollout_summaries.jsonl`: every training conversation with its reward, and `train.html` to browse them
58
+ - `logs.log`: the trainer's console log
59
+
60
+ The top level also has each run's console output (`*.out`) and the evaluation results (`run*-evals.jsonl`, `terms_sonnet_series.csv`, `identity_per_step.json`).
61
+
62
+ ## Limitations
63
+
64
+ - It confabulates freely: invented people, books and quotes are common.
65
+ - It drifts toward abstract, metaphysical, poetry-about-poetry talk. Within about 70 steps most openers had become some variation of "Do you ever write poetry?".
66
+ - Its persona is unstable: mostly a disembodied writer, occasionally a human with a life, occasionally claiming to be an AI.
67
+ - Turns were capped at 160 tokens in training, so it tends to stop mid-thought on long replies.
68
+ - It was trained only against Claude Sonnet 5.5's preferences, inside Claude Code's environment, and has not been evaluated for anything else.